{
  "source_file": "0930",
  "source_platform": "研究机构",
  "author": "Anthropic Research",
  "title_listed": "Towards Monosemanticity: Decomposing Language Models",
  "keywords": "单义性,分解,可解释性",
  "domain": "大模型算法",
  "content_type": "深度技术解析",
  "is_paid_flag": "否",
  "url": "https://transformer-circuits.pub/2023/monosemantic-features/",
  "publish_date": "2023-10",
  "corpus_stage": "预训练/SFT",
  "notes": "走向单义性：分解语言模型",
  "id": "37d897fcb626",
  "collected_at": "2026-10-02T15:31:46",
  "title": "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning",
  "body_markdown": "Transformer Circuits Thread\nTowards Monosemanticity: Decomposing Language Models With Dictionary Learning\nTowards Monosemanticity: Decomposing Language Models With Dictionary Learning\nUsing a sparse autoencoder, we extract a large number of interpretable features from a one-layer transformer.\nBrowse A/1 Features →\nBrowse All Features →\nAuthors\nTrenton Bricken\n*\n,\nAdly Templeton\n*\n,\nJoshua Batson\n*\n,\nBrian Chen\n*\n,\nAdam Jermyn\n*\n,\nTom Conerly,\nNicholas L Turner,\nCem Anil,\nCarson Denison,\nAmanda Askell,\nRobert Lasenby,\nYifan Wu,\nShauna Kravec,\nNicholas Schiefer,\nTim Maxwell,\nNicholas Joseph,\nAlex Tamkin,\nKarina Nguyen,\nBrayden McLean,\nJosiah E Burke,\nTristan Hume,\nShan Carter,\nTom Henighan,\nChris Olah\nAffiliations\nAnthropic\nPublished\nOct 4, 2023\n* Core Contributor;\nCorrespondence to\ncolah@anthropic.com\n;\nAuthor contributions statement below\n.\nMechanistic interpretability seeks to understand neural networks by breaking them into components that are more easily understood than the whole. By understanding the function of each component, and how they interact, we hope to be able to reason about the behavior of the entire network. The first step in that program is to identify the correct components to analyze.\nUnfortunately, the most natural computational unit of the neural network – the neuron itself – turns out not to be a natural unit for human understanding. This is because many neurons are\npolysemantic\n: they respond to mixtures of seemingly unrelated inputs. In the vision model\nInception v1\n, a single neuron responds to faces of cats and fronts of cars\n. In a small language model we discuss in this paper,\na single neuron\nresponds to a mixture of academic citations, English dialogue, HTTP requests, and Korean text. Polysemanticity makes it difficult to reason about the behavior of the network in terms of the activity of individual neurons.\nOne potential cause of polysemanticity is\nsuperposition\n, a hypothesized phenomenon where a neural network represents more independent \"features\" of the data than it has neurons by assigning each feature its own linear combination of neurons. If we view each feature as a vector over the neurons, then the set of features form an overcomplete linear basis for the activations of the network neurons. In our previous paper on\nToy Models of Superposition\n, we showed that superposition can arise naturally during the course of neural network training if the set of features useful to a model are sparse in the training data. As in compressed sensing, sparsity allows a model to disambiguate which combination of features produced any given activation vector.\nFor more discussion of this point, see\nDistributed Representations: Composition and Superposition\n.\nIn\nToy Models of Superposition\n, we described\nthree strategies\nto finding a sparse and interpretable set of features if they are indeed hidden by superposition: (1) creating models without superposition, perhaps by encouraging activation sparsity; (2) using dictionary learning to find an overcomplete feature basis in a model exhibiting superposition; and (3) hybrid approaches relying on a combination of the two. Since the publication of that work, we've explored all three approaches. We eventually developed counterexamples which persuaded us that the sparse architectural approach (approach 1) was insufficient to prevent polysemanticity, and that standard dictionary learning methods (approach 2) had significant issues with overfitting.\nIn this paper, we use a weak dictionary learning algorithm called a\nsparse autoencoder\nto generate\nlearned features\nfrom a trained model that offer a more monosemantic unit of analysis than the model's neurons themselves. Our approach here builds on a significant amount of prior work, especially in using dictionary learning and related methods on neural network activations (e.g.\n), and a more general allied literature on\ndisentanglement\n. We also note interim reports\nwhich independently investigated the sparse autoencoder approach in response to\nToy Models\n, culminating in the recent manuscript of Cunningham\net al.\n.\nThe goal of this paper is to provide a detailed demonstration of a sparse autoencoder compellingly succeeding at the goals of extracting interpretable features from superposition and enabling basic circuit analysis. Concretely, we take a one-layer transformer with a 512-neuron MLP layer, and decompose the MLP activations into relatively interpretable features by training sparse autoencoders on MLP activations from 8 billion data points, with expansion factors ranging from 1× (512 features) to 256× (131,072 features). We focus our detailed interpretability analyses on the 4,096 features learned in one run we call A/1.\nThis report has four major sections. In\nProblem Setup\n, we provide motivation for our approach and describe the transformers and sparse autoencoders we train. In\nDetailed Investigations of Individual Features\n, we offer an existence proof – we make the case that several features we find are functionally specific causal units which don't correspond to neurons. In\nGlobal Analysis\n, we argue that the typical feature is interpretable and that they explain a non-trivial portion of the MLP layer. Finally, in\nPhenomenology\nwe describe several properties of our features, including\nfeature-splitting\n,\nuniversality\n, and how they can form\n\"finite state automata\"-like systems\nimplementing interesting behaviors.\nWe also provide three comprehensive visualizations of features. First, for all features from\n90 learned dictionaries\nwe present activating dataset examples and downstream logit effects. We recommend the reader begin with the\nvisualization of A/1\n. Second, we provide a\ndata-oriented view\n, showing all features active on each token of 25 texts. Finally, we coembed all 4,096 features from A/1 and all 512 features from A/0 into the plane using UMAP to allow for interactive exploration of the space of features:\nSummary of Results\nSparse Autoencoders extract relatively monosemantic features.\nWe provide\nfour\ndifferent lines of evidence:\ndetailed investigations\nfor a few features firing in specific contexts for which we can construct computational proxies,\nhuman analysis\nfor a large random sample of features,\nautomated interpretability analysis of activations\nfor all the features learned by the autoencoder, and finally\nautomated interpretability analysis of logit weights\nfor all the features. Moreover, the last three analyses show that\nmost\nlearned features are interpretable. While we do not claim that our interpretations catch all aspects of features' behaviors, by constructing metrics of interpretability consistently for features and neurons, we quantitatively show their relative interpretability.\nSparse autoencoders produce interpretable features that are effectively invisible in the neuron basis.\nWe find features (e.g.,\none firing on Hebrew script\n) which are not active in any of the top dataset examples for any of the neurons.\nSparse autoencoder features can be used to intervene on and steer transformer generation.\nFor example, activating the base64 feature we study causes the model to generate base64 text, and activating the Arabic script feature we study produces Arabic text. (See\ndiscussion\nof pinned feature sampling in\nGlobal Analysis\n.)\nSparse autoencoders produce relatively universal features.\nSparse autoencoders applied to different transformer language models produce mostly similar features, more similar to one another than they are to their own model's neurons. (See\nUniversality\n)\nFeatures appear to \"split\" as we increase autoencoder size.\nWhen we gradually increase the width of the autoencoder from 512 (the number of neurons) to over 131,000 (256×), we find features which naturally fit together into families. For example, one base64 feature in a small dictionary\nsplits into three\n, with more subtle and yet still interpretable roles, in a larger dictionary. The different size autoencoders offer different \"resolutions\" for understanding the same object. (See\nFeature Splitting\n.)\nJust 512 neurons can represent tens of thousands of features.\nDespite the MLP layer being very small, we continue to find new features as we scale the sparse autoencoder.\nFeatures connect in \"finite-state automata\"-like systems that implement complex behaviors.\nFor example, we find features that work together to generate valid HTML. (See\n\"Finite State Automata\"\n.)\nProblem Setup\nA key challenge to our agenda of reverse engineering neural networks is the curse of dimensionality: as we study ever-larger models, the volume of the latent space representing the model's internal state that we need to interpret grows exponentially. We do not currently see a way to understand, search or enumerate such a space unless it can be decomposed into independent components, each of which we can understand on its own.\nIn certain limited cases, it is possible to side step these issues by rewriting neural networks in ways that don't make reference to certain hidden states. For example, in\nA Mathematical Framework for Transformer Circuits\n, we were able to analyze a one-layer attention-only network without addressing this problem. But this approach becomes impossible if we consider even a simple standard one-layer transformer with an MLP layer with a ReLU activation function. Understanding such a model requires us to have a way to decompose the MLP layer.\nIn some sense, this is the simplest language model we\nprofoundly don't understand\n. And so it makes a natural target for our paper. We aim to take its MLP activations – the activations we can't avoid needing to decompose – and decompose them into \"features\":\nCrucially, we decompose into\nmore features than there are neurons\n. This is because we believe that the MLP layer likely uses superposition\nto represent more features than it has neurons (and correspondingly do more useful non-linear work!). In fact, in our largest experiments we'll expand to have 256 times more features than neurons, although we'll primarily focus on a more modest 8× expansion.\nTransformer\nSparse Autoencoder\nLayers\n1 Attention Block\n1 MLP Block (ReLU)\n1 ReLU (up)\n1 Linear (down)\nMLP Size\n512\n512 (1×) – 131,072 (256×)\nDataset\nThe Pile\n(100 billion tokens)\nTransformer MLP Activations\n(8 billion samples)\nLoss\nAutoregressive Log-Likelihood\nL2 reconstruction\n+ L1 on hidden layer activation\nIn the following subsections, we will motivate this setup at more length. Additionally, a more detailed discussion of the architectural details and training of these models can be found in the appendix.\nFeatures as a Decomposition\nThere is significant empirical evidence suggesting that neural networks have interpretable linear directions in activation space. This includes classic work by Mikolov\net al.\ninvestigating \"vector arithmetic\" in word embeddings (\nbut see\n) and similar results in other latent spaces (e.g.\n). There is also a large body of work investigating interpretable neurons, which are just basis dimensions (e.g.\nin RNNs\n;\nin CNNs\n;\nin GANs\n;\nbut see\n). A longer review and discussion of this work can be found in the\nMotivation section\nof\nToy Models\n, although it doesn't cover more recent work (e.g.\n).\nWe'd particularly highlight an exciting exchange between Li\net al.\nand Nanda\net al.\n: Li\net al.\nfound an apparent counterexample where features were not represented as directions, which was then resolved by Nanda finding an alternative interpretation in which features were directions.\nIf linear directions are interpretable, it's natural to think there's some \"basic set\" of meaningful directions which more complex directions can be created from. We call these directions\nfeatures\n, and they're what we'd like to decompose models into. Sometimes, by happy circumstances, individual neurons appear to be these basic interpretable units (see examples above). But quite often, this isn't the case.\nInstead, we decompose the activation vector\n\\mathbf{x}^j\nas a combination of more general features which can be any direction:\n\\mathbf{x}^j \\approx \\mathbf{b} + \\sum_i f_i(\\mathbf{x}^j) \\mathbf{d}_i\n(1)\nwhere\n\\mathbf{x}^j\nis the activation vector of length\nd_{\\rm MLP}\nfor datapoint\nj\n,\nf_i(\\mathbf{x}^j)\nis the\nactivation\nof feature\ni\n, each\n\\mathbf{d}_i\nis a unit vector in activation space we call the\ndirection\nof feature\ni\n, and\n\\mathbf{b}\nis a bias.\nWhile not the focus of this paper, we could also imagine decomposing other portions of the model’s latent state or activations in this way. For instance, we could apply this decomposition to the residual stream (see e.g.\n), or to the keys and queries of attention heads, or to their outputs.\nNote that this decomposition is not new: it is just a linear matrix factorization of the kind commonly employed in dictionary learning.\nIn our sparse autoencoder setup, the feature activations are the output of the encoder\nf_i(x) = \\text{ReLU}( W_e (\\mathbf{x} - \\mathbf{b}_d) + \\mathbf{b}_e )_i,\nwhere\nW_e\nis the weight matrix of the encoder and\n\\mathbf{b}_d\n,\n\\mathbf{b}_e\nare a pre-encoder and an encoder bias. The feature directions are the columns of the decoder weight matrix\nW_d\n. (See\nappendix\nfor full notation.)\nIf such a sparse decomposition exists, it raises an important question: are models in some fundamental sense composed of features or are features just a convenient post-hoc description? In this paper, we take an agnostic position, though our\nresults\non feature universality suggest that features have some existence beyond individual models.\nSuperposition Hypothesis\nTo see how this decomposition relates to superposition, recall that the superposition hypothesis postulates that neural networks “want to represent more features than they have neurons”. We think this happens via a kind of “noisy simulation”, where small neural networks exploit feature sparsity and properties of high-dimensional spaces to approximately simulate much larger much sparser neural networks\n.\nA consequence of this is that we should expect the feature directions to form an\novercomplete basis\n. That is, our decomposition should have more directions\n\\mathbf{d}_i\nthan neurons. Moreover, the feature activations should be\nsparse\n, because sparsity is what enables this kind of noisy simulation. This is mathematically identical to the classic problem of dictionary learning.\nWhat makes a good decomposition?\nSuppose that a dictionary exists such that the MLP activation of each datapoint is in fact well approximated by a sparse weighted sum of features as in equation 1. That decomposition will be useful for interpreting the neural network if:\nWe can interpret the conditions under which each feature is active. That is, we have a description of which datapoints\nj\ncause the feature to activate (i.e. for which\nf_i(\\mathbf{x}^j)\nis large) that makes sense on a diverse dataset (including potentially synthetic or off-distribution examples meant to test the hypothesis).\nThis property is closely related to the desiderata of Causality, Generality, and Purity discussed in Cammarata\net al.\n, and those provide an example of how we might make this property concrete in a specific instance.\nWe can interpret the downstream effects of each feature, i.e. the effect of changing the value of\nf_i\non subsequent layers. This should be consistent with the interpretation in (1).\nThe features explain a significant portion of the functionality of the MLP layer (as measured by the loss; see\nHow much of the model does our interpretation explain\n?).\nA feature decomposition satisfying these criteria would allow us to:\nDetermine the contribution of a feature to the layer’s output, and the next layer’s activations, on a specific example.\nMonitor the network for the activation of a specific feature (see e.g. speculation about\nsafety-relevant features\n).\nChange the behavior of the network in predictable ways by changing the values of some features. In multilayer models this could look like predictably influencing one layer by changing the feature activations in an earlier layer.\nDemonstrate that the network has learned certain properties of the data.\nDemonstrate that the network is using a given property of the data in producing its output on a specific example.\nThis is similar in spirit to the evidence provided by influence functions.\nDesign inputs meant to activate a given feature and elicit certain outputs.\nOf course, decomposing models into components is just the beginning of the work of mechanistic interpretability! It provides a foothold on the inner workings of models, allowing us to start in earnest on the task of unraveling circuits and building a larger-scale understanding of models.\nWhy not use architectural approaches?\nIn\nToy Models of Superposition\n, we highlighted several different approaches to solving superposition. One of those approaches was to engineer models to simply not have superposition in the first place.\nInitially, we thought that this might be possible but come with a large performance hit (i.e. produce models with greater loss). Even if this performance hit had been too large to use in practice for real models, we felt that success at creating monosemantic models would have been very useful for research, and in a lot of ways this felt like the \"cleanest\" approach for downstream analysis.\nUnfortunately, having spent a significant amount of time investigating this approach, we have ultimately concluded that it is more fundamentally non-viable.\nIn particular, we made several attempts to induce activation sparsity during training to produce models without superposition, even to the point of training models with 1-hot activations. This indeed eliminates superposition, but it fails to result in cleanly-interpretable neurons! Specifically, we found that individual neurons can be polysemantic even in the absence of superposition. This is because in many cases models achieve lower loss by representing multiple features ambiguously (in a polysemantic neuron) than by representing a single feature unambiguously and ignoring the others.\nTo understand this, consider a toy model with a single neuron trained on a dataset with four mutually-exclusive features (A/B/C/D), each of which makes a distinct (correct) prediction for the next token, labeled in the same fashion. Further suppose that this neuron’s output is binary: it either fires or it doesn’t. When it fires, it produces an output vector representing the probabilities of the different possible next tokens.\nWe can calculate the cross-entropy loss achieved by this model in a few cases:\nSuppose the neuron only fires on feature A, and correctly predicts token A when it does. The model ignores all of the other features, predicting a uniform distribution over tokens B/C/D when feature A is not present. In this case the loss is\n\\frac{3}{4}\\ln 3 \\approx 0.8\n.\nInstead suppose that the neuron fires on both features A and B, predicting a uniform distribution over the A and B tokens. When the A and B features are not present, the model predicts a uniform distribution over the C and D tokens. In this case the loss is\n\\ln 2 \\approx 0.7\n.\nBecause the loss is lower in case (2) than in case (1), the model achieves better performance by making its sole neuron polysemantic, even though there is no superposition.\nThis example might initially seem uninteresting because it only involves one neuron, but it actually points at a general issue with highly sparse networks. If we push activation sparsity to its limit, only a single neuron will activate at a time. We can now consider that single neuron and the cases where it fires. As seen earlier, it can still be advantageous for that neuron to be polysemantic.\nBased on this reasoning, and the results of our experiments, we believe that models trained on cross-entropy loss will generally prefer to represent more features polysemantically than to represent fewer \"true features\" monosemantically, even in cases where sparsity constraints make superposition impossible.\nModels trained on other loss functions do not necessarily suffer this problem. For instance, models trained under mean squared error loss (MSE) may achieve the same loss for both polysemantic and monosemantic representations (e.g.\n), and for some feature importance curves they may actively prefer the monosemantic solution\n. But language models are not trained with MSE, and so we do not believe architectural changes can be used to create fully monosemantic language models.\nNote, however, that in learning to decompose models post-training we do use an MSE loss (between the activations and their representation in terms of the dictionary), so sparsity can inhibit superposition from forming in the learned dictionary. (Otherwise, we might have superposition \"all the way down.\")\nUsing Sparse Autoencoders To Find Good Decompositions\nThere is a long-standing hypothesis that many natural latent variables in the world are sparse (\nsee\n, section \"Why Sparseness?\"). Although we have a limited understanding of the features that exist in language models, the examples we do have (e.g.\n) are suggestive of highly sparse variables. Our work on Toy Models of Superposition\nshows that a large set of sparse features could be represented in terms of directions in a lower dimensional space.\nFor this reason, we seek a decomposition which is\nsparse\nand\novercomplete\n. This is essentially the problem of\nsparse dictionary learning\n. (Note that, since we're only asking for a sparse decomposition of the activations, we make no use of the downstream effects of\n\\mathbf{d}_i\non the model's output. This means we'll be able to use those downstream effects in later sections as a kind of validation on the features found.)\nIt's important to understand why making the problem\novercomplete\n– which might initially sound like a trivial change – actually makes this setting very different from similar approaches seeking sparse disentanglement in the literature. It's closely connected to why dictionary learning is such a non-trivial operation; in fact, as we'll see, it's actually kind of miraculous that this is possible at all. At the heart of dictionary learning is an\ninner problem\nof computing the feature activations\nf_i(\\mathbf{x})\nfor each datapoint\n\\mathbf{x}\n, given the feature directions\n\\mathbf{d}_i\n. On its surface, this problem may seem impossible: we're asking to determine a high-dimensional vector from a low-dimensional projection. Put another way, we're trying to invert a very rectangular matrix. The only thing which makes it possible is that we are looking for a high-dimensional vector that is sparse! This is the famous and well-studied problem of\ncompressed sensing\n, which is NP-hard in its exact form. It is possible to store high-dimensional sparse structure in lower-dimensional spaces, but recovering it is hard.\nDespite its difficulty, there are a host of sophisticated methods for dictionary learning (e.g.\n). These methods typically involve optimizing a relaxed problem or doing a greedy search. We tried several of these classic methods, but ultimately decided to focus on a simpler sparse autoencoder approximation of dictionary learning (similar to Sharkey,\net al.\n). This was for two reasons. First, a sparse autoencoder can readily scale to very large datasets, which we believe is necessary to characterize the features present in a model trained on a large and diverse corpus.\nFor the model discussed in this paper, we trained the autoencoder on 8 billion datapoints.\nSecondly, we have a concern that iterative dictionary learning methods might be\n\"too strong\"\n, in the sense of being able to recover features from the activations which the model itself cannot access. Exact compressed sensing is NP-hard, which the neural network certainly isn't doing. By contrast, a sparse autoencoder is very similar in architecture to the MLP layers in language models, and so should be similarly powerful in its ability to recover features from superposition.\nSparse Autoencoder Setup\nWe briefly overview the architecture and training of our sparse autoencoder here, and provide further details in\nBasic Autoencoder Training\n. Our sparse autoencoder is a model with a bias at the input, a linear layer with bias and ReLU for the encoder, and then another linear layer and bias for the decoder. In toy models we found that the bias terms were quite important to the autoencoder’s performance.\nWe tie the biases applied in the input and output, so the result is equivalent to subtracting a fixed bias from all activations and then using an autoencoder whose only bias is before the encoder activation.\nWe train this autoencoder using the Adam optimizer to reconstruct the MLP activations of our transformer model, with an MSE\nNote that using an MSE loss avoids the challenges with polysemanticity that we discussed above in\nWhy Not Architectural Approaches?\nloss plus an L1 penalty to encourage sparsity.\nIn training the autoencoder, we found a couple of principles to be quite important. First, scale really matters. We found that training the autoencoder on more data made features subjectively “sharper” and more interpretable. In the end, we decided to use 8 billion training points for the autoencoder (see\nAutoencoder Dataset\n).\nSecond, we found that over the course of training some neurons cease to activate, even across a large number of datapoints. We found that “resampling” these dead neurons during training gave better results by allowing the model to represent more features for a given autoencoder hidden layer dimension. Our resampling procedure is detailed in\nNeuron Resampling\n, but in brief we periodically check for neurons which have not fired in a significant number of steps and reset the encoder weights on the dead neurons to match data points that the autoencoder does not currently represent well.\nFor readers looking to apply this approach, we supply an appendix with\nAdvice for Training Sparse Autoencoders\n.\nHow can we tell if the autoencoder is working?\nUsually in machine learning we can quite easily tell if a method is working by looking at an easily-measured quantity like the test loss. We spent quite some time searching for an equivalent metric to guide our efforts here, and unfortunately have yet to find anything satisfactory.\nWe began by looking for\nan information-based metric\n, so that we could say in some sense that the best factorization is the one that minimizes the total information of the autoencoder and the data. Unfortunately, this total information did not generally correlate with subjective feature interpretability or activation sparsity. (Runs whose feature activations had an average L0 norm in the hundreds but low reconstruction error could have lower total information than those with smaller average L0 norm and higher reconstruction error.)\nThus we ended up using a combination of several additional metrics to guide our investigations:\nManual inspection:\nDo the features seem interpretable?\nFeature density\n:\nwe found that the number of “live” features and the percentage of tokens on which they fire to be an extremely useful guide. (See appendix for\ndetails\n.)\nReconstruction loss:\nHow well does the autoencoder reconstruct the MLP activations? Our goal is ultimately to explain the function of the MLP layer, so the MSE loss should be low.\nToy models:\nHaving toy models where we know the ground truth and so can cleanly evaluate the autoencoder’s performance was crucial to our early progress.\nInterpreting or measuring some of these signals can be difficult, though. For instance, at various points we thought we saw features which at first didn’t make any sense, but with deeper inspection we\ncould understand\n. Likewise, while we have identified some desiderata for the distribution of feature densities, there is much that we still do not understand and which prevents this from providing a clear signal of progress.\nWe think it would be very helpful if we could identify better metrics for dictionary learning solutions from sparse autoencoders trained on transformers.\nThe (one-layer) model we're studying\nWe chose to study a one-layer transformer model. We view this model as a testbed for dictionary learning, and in that role it brings three key advantages:\nBecause one-layer models are small they likely have fewer \"true features\" than larger models, meaning features learned by smaller dictionaries might cover all their \"true features\". Smaller dictionaries are cheaper to train, allowing for fast hyperparameter optimization and experimentation.\nWe can highly overtrain a one-layer transformer quite cheaply. We hypothesize that a very high number of training tokens may allow our model to learn cleaner representations in superposition.\nWe can easily analyze the effects features have on the logit outputs, because these are approximately linear in the feature activations.\nNote that this linear structure makes it even more likely that features should be linear. On the one hand, this means that the linear representation hypothesis is more likely to hold for this model. On the other hand, it potentially means that our results are less likely to generalize to multilayer models. Fortunately, others have studied multilayer transformers with sparse autoencoders and found interpretable linear features, which gives us more confidence that what we see in the one-layer model indeed generalizes\n.\nThis provides helpful corroboration that the features we have found are not just telling us about the data distribution, and actually\nreflect the functionality of the model\n.\nWe trained two one-layer transformers with the same hyperparameters and datasets, differing only in the random seed used for initialization. We then learned dictionaries of many different sizes on both transformers, using the same hyperparameters for each matched pair of dictionaries but training on the activations of different tokens for each transformer.\nWe refer to the main transformer we study in this paper as the “A” transformer. We primarily use the other transformer (“B”) to study\nfeature universality\n, as we can e.g. compare features learned from the “A” and “B” transformers and see how similar they are.\nNotation for Features\nThroughout this draft, we'll use strings like \"A/1/2357\" to denote features. The first portion \"A\" or \"B\" denote which model the features come from. The second part (e.g. the \"1\" in \"A/1\") denotes the dictionary learning run. These vary in the number of learned factors and the L1 coefficient used. A table of all of our runs is available\nhere\n. Notably, A/0…A/5 form a sequence with fixed L1 coefficients and increasing dictionary sizes. The final portion (e.g. the \"2357\" in \"A/1/2357\") corresponds to the specific feature in the run.\nSometimes, we want to denote neurons from the transformer rather than features learned by the sparse autoencoder. In this case, we use the notation \"A/neurons/32\".\nInterface for Exploring Features\nWe provide an interface for exploring all the features in all our dictionary learning runs. Links to the visualizations for each run can be found\nhere\n. We suggest beginning with the\ninterface for A/1\n, which we discuss the most.\nThese interfaces provide extensive information on each feature. This includes examples of when they activate, what effect they have on the logits when they do, examples of how they affect the probability of tokens if the feature is ablated, and much more:\nOur interface also allows users to search through features:\nAdditionally, we provide a\nsecond\ninterface\ndisplaying all features active on a given dataset example. This is available for a set of example texts.\nDetailed Investigations of Individual Features\nThe most important claim of our paper is that dictionary learning can extract features that are significantly more monosemantic than neurons. In this section, we give a detailed demonstration of this claim for a small number of features which activate in highly specific contexts.\nThe features we study respond to\nText written in Arabic script (like \"الكتاب المختصر في حساب الجبر والمقابلة\")\nDNA sequences (like \"CCTGGTACTGTACGAACGAACGAACGTAGCCTTGG\")\nbase64 strings (like the final characters in \"https://www.youtube.com/watch?v=\ndQw4w9WgXcQ\n\")\nText written in Hebrew script (like \"בראשית ברא אלהים את השמים ואת הארץ\")\nFor each learned feature, we attempt to establish the following claims:\nThe learned feature\nactivates with high specificity\nfor the hypothesized context. (When the feature is on the context is usually present.)\nThe learned feature\nactivates with high sensitivity\nfor the hypothesized context. (When the context is present, the feature is usually on.)\nThe learned feature\ncauses\nappropriate downstream behavior.\nThe learned feature\ndoes not correspond to any neuron\n.\nThe learned feature is\nuniversal\n– a similar feature is found by dictionary learning applied to a different model.\nTo demonstrate claims 1–3, we devise computational proxies for each context, numerical scores estimating the (log-)likelihood that a string (or token) is from the specific context. The contexts chosen above are easy to model based on the defined sets of unicode characters involved. We model DNA sequences as random strings of characters from\n[ATCG]\nand we model base64 strings as random sequences of characters from\n[a-zA-Z0-9+/]\n. For Arabic script and Hebrew features, we exploit the fact that each language is written in a script consisting of well-defined Unicode blocks. Each computational proxy is then an estimate of the log-likelihood ratio of a string under the hypothesis versus under the full empirical distribution of the dataset. The full description of how we estimate\n\\log(P(s|\\text{context}) / P(s))\nfor each feature hypothesis is given in the\nappendix section on proxies\n.\nIn this section we primarily study the learned feature which is most active in each context. There are typically other features that also model that context, and we find that rare “gaps” in the sensitivity of a main feature are often explained by the activity of another. We discuss this phenomenon in detail in sections on\nActivation Sensitivity\nand\nFeature Splitting\n.\nWe take pains to demonstrate the specificity of each feature, as we believe that to be more important for ruling out polysemanticity. Polysemanticity typically involves neurons activating for clearly unrelated concepts.\nFor example,\none neuron\nin our transformer model responds to a mix of academic citations, English dialogue, HTTP requests, and Korean text. In vision models, there is a classic example of a neuron which responds to cat faces and fronts of cars\n.\nIf our proxy shows that a feature only activates in some relatively rare and specific context, that would exclude the typical form of polysemanticity.\nWe finally note that the features in this section are cherry-picked to be\neasier to analyze\n. Defining simple computational proxies for most features we find, such as text concerning\nfantasy games\n, would be difficult, and we analyze them in other ways in the\nfollowing section\n.\nArabic Script Feature\nThe first feature we'll consider is an Arabic Script feature,\nA/1/3450\n. It activates in response to text in Arabic, Farsi, Urdu (and possibly other languages), which use the Arabic script. This feature is quite specific and relatively sensitive to Arabic script, and effectively invisible if we view the model in terms of individual neurons.\nActivation Specificity\nOur first step is to show that this feature fires almost exclusively on text in Arabic script. We give each token an \"Arabic script\" score using an estimated likelihood ratio\n\\log(P(s|\\text{Arabic Script}) / P(s))\n, and break down the histogram of feature activations by that score. Arabic text is quite rare in our overall data distribution –  just 0.13% of training tokens — but it makes up 81% of the tokens on which our feature is active. That percentage varies significantly by feature activity level, from 25% when the feature is barely active to 98% when the feature activation is above 5.\nWe also show dataset examples demonstrating different levels of feature activity. In interpreting them, it's important to note that Arabic Unicode characters are often split into multiple tokens. For example, the character\nث\n(\nU+062B\n) is tokenized as\n\\xd8\nfollowed by\n\\xab\n.\nThis kind of split tokenization is common for Unicode characters outside the Latin script and the most common characters from other Unicode blocks.\nWhen this happens, A/1/3450 typically only fires on the\nlast\ntoken of the Unicode character.\nThe upper parts of the activation spectrum, above an activity of ~5, clearly respond with high specificity to Arabic script. What should we make of the lower portions? We have three hypotheses:\nThe proxy is imperfect.\nIt has false negatives due to common characters which are in other Unicode blocks (e.g., whitespace, punctuation), and also due to places where tokenization has created strange behavior.\nFor example, in the figure above, there is a newline character ⏎ that our feature fires on but our proxy assigns a low score to because it is outside the Unicode block.\nThe model may be imperfect (but still calibrated).\nIf features activate proportional to their \"confidence\" that some property is present, then we should expect them to sometimes be wrong in their weak activations. Although it might seem silly to fail to recognize which language a piece of text is, this is a very weak one layer transformer model, and the fact that characters are often split into multiple tokens adds additional difficulty.\nThe autoencoder may be imperfect:\nIf the width of our autoencoder is less than the number of \"true features\" being used by the model, then unrecovered features may show up as low activations across many of our learned features.\nRegardless, large feature activations have larger impacts on model predictions,\nWhy do we believe large activations have larger effects? In the case of a one-layer transformer like the one we consider in this paper, we can make a strong case for this: features have a linear effect on the logits (modulo rescaling by layer norm), and so a larger activation of the feature has a larger effect on the logits. In the case of larger models, this follows from a lot of conjectures and heuristic arguments (e.g. the abundance of linear pathways in the model and the idea of linear features at each layer), and must be true for sufficiently small activations by continuity, but doesn't have a watertight argument.\nso getting their interpretation right matters most. One useful tool for assessing the impact of false positives at low activation levels is the \"expected value plot\" of Cammarata\net al.\n. We plot the distribution of feature activations\nweighted by activation level\n. Most of the magnitude of activation provided by this feature comes from dataset examples which are in Arabic script.\nActivation Sensitivity\nIn the Feature Activation Distribution above, it's clear that A/1/3450 is not sensitive to all tokens in Arabic script. In the random dataset examples, it fails to fire on five examples of the prefix \"ال\", transliterated as \"al-\", which is the equivalent of the definite article \"the\" in English. However, in exactly those places, another feature which is specific to Arabic script,\nA/1/3134\n, fires. There are several additional features that fire on Arabic and related scripts (e.g.\nA/1/1466\n,\nA/1/3134\n,\nA/1/3399\n) which contribute to representing Arabic script. Another example deals with Unicode tokenization: when Arabic characters are split into multiple tokens, the feature we analyze here only activates at the final token comprising the character, while\nA/1/3399\nactivates on the first token comprising the character.  To see how these features collaborate, we provide an\nalternative visualization\nshowing all the features active on a snippet of Arabic text. We consider such interactions more in the\nPhenomenology\nsection below.\nNevertheless, we find a Pearson correlation of 0.74 between the activity of our feature and the activity of the Arabic script proxy (thresholded at 0), over a dataset of 40 million tokens. Correlation provides a joint measure of sensitivity and specificity that takes magnitude into account, and 0.74 is a substantial correlation.\nFeature Downstream Effects\nBecause the autoencoder is trained on model activations, the features it learns could in theory represent structure in the training data alone, without any relevance to the network’s function. We show instead that the learned features have interpretable causal effects on model outputs which make sense in light of the features’ activations. Note that these downstream effects are not inputs to the dictionary learning process, which only sees the activations of the MLP layer. If the resulting features also mediate important downstream\nbehavioral effects\nthen we can be confident that the feature is truly connected to the MLP’s functional role in the network and not just a property of the underlying data.\nWe begin with a linear approximation to the effect of each feature on the model logits. We compute the logit weight following the path expansion approach of\n,\n(also called “logit attribution”, see similar work e.g.\n)\nmultiplying each feature direction by the MLP output weights, an approximation of the layer norm operation combining a projection to remove the mean and a diagonal matrix approximating the scaling terms, and the unembedding matrix (\nd_iW_{down}\\pi L W_{unembed}\n). Because the softmax function is shift-invariant there is no absolute scale for the logit weights; we shift these so that the median logit weight for each feature is zero.\nEach feature, when active, makes some output tokens more likely and some output tokens less likely. We plot that distribution\nof logit weights.\nWe exclude weights corresponding to extremely rare or never used vocabulary elements. These are perhaps similar to the \"anomalous tokens\" (e.g., \"SolidGoldMagikarp\") of Rumbelow & Watkins\n.\nThere is a large primary mode at zero, and a second much smaller mode on the far right, the tokens whose likelihood most increases when our feature is on. This second mode appears to correspond to Arabic characters, and tokens which help represent Arabic script characters (especially\n\\xd8\nand\n\\xd9\n, which are often the first half of the UTF-8 encodings of Arabic Unicode characters in the\nbasic Arabic Unicode block\n).\nThis suggests that activating this feature increases the probability the network predicts Arabic script tokens.\nThere are in theory several ways the logit weights could overestimate the model's actual use of a feature:\n1. It could be that these output weights are small enough that, when multiplied by activations, they don't have an appreciable effect on the model’s output.\n2. The feature might only activate in situations where other features make these tokens extremely unlikely, such that the feature in fact has little effect.\n3. It is possible that our approximation of linearizing the layer norm (see\nFramework\n) is poor.\nBased on the subsequent analysis, which confirms the logit weight effects, we do not believe these issues arise in practice.\nTo visualize these effects on actual data, we\ncausally\nablate\nthe feature\n. For a given dataset example, we run the context through the model until the MLP layer, decode the activations into features, then subtract off the activation of\nA/1/3450\n, artificially setting it to zero on the whole context, before applying the rest of the model. We visualize the effect of ablating the feature using underlines in the visualization; tokens whose predictions were helped by the feature (ablation decreased likelihood) are underlined in blue and tokens whose predictions were hurt by the feature (ablation increased likelihood) are underlined in red.\nIn the example on the right below we see that the\nA/1/3450\nwas active on every token in a short context (orange background). Ablating it hurt the predictions of all the tokens in Arabic script (purple underlines), but helped the prediction of the period\n.\n(orange underline). The rest of the figure displays contexts from two different ranges of feature activation levels. (The feature activation on the middle token of examples on the right (\"subsample interval 5\") is about half that of the middle token of examples on the left (\"subsample interval 0\")). We see that the feature was causally helping the model predictions on Arabic script through that full range, and the only tokens made less likely by the feature are punctuation shared with other scripts. The magnitudes of the impact are larger when the feature is more active.\nWe encourage interested readers to view the feature visualization for\nA/1\nto review this and other effects.\nWe also validate that the feature's downstream effect is in line with our interpretation as an Arabic script feature by sampling from the model with the feature activity \"pinned\" at a high value. To do this, we start with a prefix\n1,2,3,4,5,6,7,8,9,10\nwhere the model has an expected continuation (keep in mind that this is a one layer model that is very weak!). We then instead set\nA/1/3450\nto its maximum observed value and see how that changes the samples:\nThe feature is not a neuron\nThis feature seems rather monosemantic, but some models have relatively monosemantic neurons, and we want to check that dictionary learning didn't merely hand us a particularly nice neuron.\nOf course, some features might genuinely be neuron aligned. But we'd like to know that at least some of the features dictionary learning discovers were not trivial.\nWe first note that when we\nsearch\nfor neurons with Arabic script in their top dataset 20 examples, only one neuron turns up, with a single example in Arabic script. (Eighteen of the remaining examples are in English, and one is in Cyrillic.)\nWe then look at the coefficients of the feature in the neuron basis, and find that the three largest coefficients by magnitude are all negative (!) and there are a full 27 neurons whose coefficients are at least 0.1 in magnitude.\nIt is of course possible that these neurons engage in a delicate game of cancellation, resulting in one particular neuron's primary activations being sharpened. To check for this, we find the neuron whose activations are most correlated to the feature's activations over a set of ~40 million dataset examples.\nWe also tried looking at the neurons which have the largest contribution to the dictionary vector for the feature. However, we found looking at the most correlated neuron to be more reliable – in earlier experiments, we found rare cases where the correlated method found a seemingly similar neuron, while the dictionary method did not. This may be because neurons can have different scales of activations.\nThe most correlated neuron (\nA/neurons/489\n) responds to a mixture of different non-English languages.\nThis makes sense as a superposition strategy: since languages are essentially mutually exclusive, they're natural to put in superposition with each other\nWe can see this by visualizing the activation distribution of the neuron – the dataset examples are all other languages, with Arabic script present only as a small sliver.\nLogit weight analysis is also consistent with this neuron responding to a mixture of languages. For example, in the figure below many of the top logit weights appear to include Russian and Korean tokens. Careful readers will observe a thin red sliver corresponding to rare Arabic script tokens in the distribution. These Arabic script tokens have weight values that are very slightly positive leaning overall, but some are negative.\nFinally, scatter plots and correlations suggest the similarities between\nA/1/3450\nand the neuron are non-zero, but quite minimal.\nBy analogy, this also applies to\nB/1/1334\n.\nIn particular, note how the y-axis of the logit weights scatter plot (corresponding to the feature) cleanly separates the Arabic script token logits, while the x-axis does not. More generally, note how the y-marginals both clearly exhibit specificity, but the x-marginals do not.\nWe conclude that the features we study do not trivially correspond to a single neuron. The Arabic script feature would be effectively invisible if we only analyzed the model in terms of neurons.\nUniversality\nWe will now ask whether A/1/3450 is a\nuniversal\nfeature that forms in other models and can be consistently discovered by dictionary learning. This would indicate we are discovering something more general about how one-layer transformers learn representations of the dataset.\nWe search for a similar feature in\nB/1\n, a dictionary learning run on a transformer trained on the same dataset but with a different random seed. We search for the feature with the highest activation correlation\n\"Activation correlation\" is defined as the feature whose activations across 40,960,000 tokens has the highest Pearson correlation with those of\nA/1/3450\n.\nand find\nB/1/1334\n(corr=0.91), which is strikingly similar:\nThis feature clearly responds to Arabic script as well. If anything, it's nicer than our original feature – it's more specific in the 0–1 range. The logit weights tell a similar story:\nThe effects of ablating this feature are also consistent with this (see the visualization for\nB/1/1334\n).\nTo more systematically analyze the similarities between A and B, we look at scatter plots comparing the activations or logit weights:\nThe activations are strongly correlated (Pearson correlation of 0.91), especially in the main Arabic mode.\nThe logit weights reveal a two-dimensional version of the bimodality we saw in the histogram for A/1, with logit weights for Arabic tokens clustering at the top right. The correlation is more modest than that of the activations because the distribution is dominated by a relatively uncorrelated mode in the center. We hypothesize this central mode corresponds to \"weight interference\" and that the shared outlier mode is the important observation – that is, the model may ideally prefer to have all those weights be zero, but due to superposition with other features and their weights, this isn't possible.\nDNA Feature\nWe now consider a DNA feature,\nA/1/2937\n. It activates in response to long uppercase strings consisting of\nA\n,\nT\n,\nC\n, and\nG\n, typically used to represent nucleotide sequences. We closely follow the analysis of the Arabic script feature above to show activation specificity and sensitivity for the feature, sensible downstream effects, a lack of neuron alignment, and universality between models. The main differences will be that (1)\nA/1/2937\nis the only feature devoted to modeling the DNA context, and (2) our proxy is less sensitive to DNA than our feature is, missing strings containing punctuation, spaces, and missing bases.\nActivation Specificity and Sensitivity\nWe begin with the computational proxy for \"is a DNA sequence\",\n\\log(P(s|\\text{DNA}) / P(s))\n. Because there are some vocabulary tokens, such as\nCAT\n, which could occur in DNA but also occur in other contexts, we always look at groups of at least two tokens when evaluating the proxy. The log-probabilities turn out to be quite bimodal, so we binarize the proxy (based on its sign). This binarized proxy then has a Pearson correlation of 0.8 with the feature activations.\nWhile the feature appears to be quite monosemantic in the feature's higher registers (all 10 random dataset examples above activation of 6.0 are DNA sequences), there is significant blue indicating the DNA proxy not firing in the lower registers. Below we show a grid of random examples at four activation levels (including feature off) where the proxy does and doesn't fire.\nWe note that in all but two cases where the feature and proxy disagree, the feature does indicate a DNA sequence, just one outside our proxy's strict\nATCG\nvocabulary. For example, the space present in the triplets\nTGG AGT\nmakes the proxy fail to fire. The feature also fires productively on\n'-\nin the string\n5'-TCT\n, because what follows that prefix should be DNA, even though the prefix is not itself DNA. (The causal ablation reveals that turning off the DNA feature hurts the prediction of the strings that follow.) We thus believe that\nA/1/2937\nis quite sensitive and specific for DNA. The case where the proxy fires and the feature does not is indeed a DNA sequence, though the feature begins firing on the very next token. We observe that the DNA feature may not fire on the first few tokens of a DNA sequence,\nbut by the end of a long DNA sequence\n, it is the\nonly\nfeature active.\nFeature Downstream Effect\nThe downstream effect of the DNA feature being active, as measured by logit weights, make sense, with all the top tokens being combinations of nucleotides like\nAGT\nand\nGCC\n.\nThe Feature Is Not A Neuron\nThe most similar neuron to\nA/1/2937\n, as measured by activation correlation, is\nA/neurons/67\n. DNA contexts form a tiny sliver of that neuron's activating examples. The neuron whose coefficient is the highest in our feature's vector,\nA/neurons/227\n, also has no DNA sequences in its top activating examples.\nUniversality\nA/1/2937\nhas a correlated feature (corr=0.92) in run B/1,\nB/1/3680\n. Their top logit weights agree, and are DNA tokens (e.g.\nAGT\n) forming a separate mode (circled on the right) than the bulk of the logit weights for both.\nBase64 Feature\nWe now consider a\nbase64\nfeature,\nA/1/2357\n. We're particularly excited by this feature because we discovered a base64 neuron in our SoLU paper\n, suggesting that base64 might be quite universal – even across models trained on somewhat different datasets and with different architectures. We model base64 strings as random sequences of characters from\n[a-zA-Z0-9+/]\n. The activation distribution colored by the corresponding computational proxy, together with the random dataset examples from each activation level, shows that this feature is quite specific to base64.\nThis is not the only feature active in base64 contexts, and in a section\nbelow\nwe discuss the two others, one of which fires on single digits in base64 contexts (like the 2, 4, 7, and 9 on which A/1/2357 doesn't activate in the figure above), exploiting a property of the BPE tokenizer to make a better prediction.\nTurning to the logit weights, they have a second mode consisting of highly base64-specific tokens. The main mode seems to primarily be interference, but the right side is skewed towards base64-neutral or slightly base64-leaning tokens. (If we look at the conditional below, we see a more continuous transition to base64-specific tokens.)\nThere is a more continuous transition between non-base64 and base64 tokens than we saw in the Arabic script example. This difference likely arises because whether a token occurs more in Arabic script than in other text is a relatively binary distinction, whereas whether a token occurs more in base64 or other text varies more continuously. For instance,\nfr\nis both a common abbreviation for the French language and also a base64 token, so it makes sense for the model to be cautious in up-weighting\nfr\nbecause it might already have a higher prior due to use in French. Indeed, any token consisting of letters from the English alphabet will have some nontrivial probability of appearing in base64 strings.\nThe Pearson correlation between the computational proxy and the activity of\nA/1/2357\nis just 0.38. We believe that is mostly because the proxy is too broad. For example hexadecimal strings (those made of\n[0-9A-F]\n) activate the proxy, as they are quite different from the overall data distribution, but are actually predicted by a feature of their own,\nA/1/3817\n.\nUniversality\nA/1/2357\nhas a correlated feature (corr=0.85) in run B/1,\nB/1/2165\n. It also has high activation specificity for base64 strings:\nLike\nA/1/2357\n,\nB/1/2165\n's logit weights have a second mode corresponding to base64 tokens:\nCorrelations and scatter plots are also consistent with them being very similar features:\nNote that we expect the overlap between the interference and base64 token logit weights to be from the aforementioned usage of base64 tokens across many other contexts.\nThe Feature Is Not A Neuron\nLooking at the neuron in model A that most correlates with this feature:\nA/neurons/470\n(corr=0.18), we find that while it does notably respond to base64 strings, it also activates for lots of other things, including code, HTML labels, parts of URLs, etc.:\nThe logit weights suggest it somewhat increases base64 tokens, but is much more focused on upweighting other tokens, e.g. filename endings.\nThe activation and logit correlations are consistent with this neuron helping represent the same feature, but largely doing other things.\nHebrew Feature\nAnother interesting example is the Hebrew feature\nA/1/416\n. Like the Arabic feature, it's easy to computationally identify Hebrew text based on Unicode blocks.\nA/1/416\nhas high activation specificity in the upper spectrum. It does weakly activate for other things (especially other languages with Unicode scripts). There is also some blue in strong activations; this appears to significantly be on \"common characters\", such as whitespace or punctuation, which are from other unicode blocks (see more discussion of similar issues in the\nArabic feature section\n).\nIts logit weights have a notable second mode, corresponding to Hebrew characters and relevant incomplete Unicode characters. Note that\n\\xd7\nis the first token in the UTF-8 encoding of most characters in the\nbasic Hebrew Unicode block\n.\nThe Pearson correlation of the Hebrew script proxy with\nA/1/416\nis 0.55. Some of the failure of sensitivity may be due to a complementary feature\nA/1/1016\nthat fires on\n\\xd7\nand predicts the bytes that complete Hebrew characters' codepoints.\nThe Feature Is Not A Neuron\nThere doesn't appear to be a similar neuron. The most correlated neuron in model A is\nA/neurons/489\n(corr=0.1), which has low activation and logit specificity. Consider the following activation and logit correlation plots:\nTo cross-validate this, we also\nsearched\nfor any neuron where the main Hebrew Unicode block appeared in the top dataset examples. We found none.\nUniversality\nA/1/416\nhas a correlated feature in the B/1 run,\nB/1/1901\n(corr=0.92) that has significant activation specificity:\nLogit weights have a second mode, as before:\nActivation and logit weight correlations are again consistent:\nGlobal Analysis\nIf the previous section has persuaded you that at least some of the features are genuinely interpretable and reflect the underlying model mechanics, it's natural to wonder how broadly this holds outside of those cherry-picked features. The primary focus of this section will be to answer the question, \"how interpretable are the rest of the features?\" We show that both humans and large language models find our features to be significantly more interpretable than neurons, and quite interpretable in absolute terms.\nThere are a number of other questions one might also ask. To what extent is our dictionary learning method discovering all the features necessary to understand the MLP layer? Holistically, how much of the MLP layer's mechanics have been made interpretable? We are not yet able to fully answer these questions to our satisfaction, but will provide some preliminary speculation towards the end of this section.\nWe note that of the 4,096 learned features in the A/1 autoencoder, 168 of them are \"dead\" (active on none of the 100 million dataset examples) and 292 of them are \"ultralow density\", active on less than 1 in a million dataset examples and exhibiting\nother atypical properties\n. We exclude both these groups of features from further analyses.\nHow Interpretable is the Typical Feature?\nIn this section, we use three different methods to analyze how interpretable the typical feature is, and how that compares to neurons: human analysis, and two forms of automated interpretability.\nAll three approaches find that features are much more interpretable than neurons.\nManual Human Analysis\nAt present, we do not have any metric we trust more than human judgment of interpretability. Thus, we had a blinded annotator (one of the authors, Adam Jermyn) score features and neurons based on how interpretable they are. The scoring rubric can be found in the\nappendix\nand accounts for confidence in an explanation, consistency of the activations with that explanation, consistency of the logit output weights with that explanation, and specificity.\nIn doing this evaluation, we wanted to avoid a weakness we perceived in our prior work (e.g.,\n) of focusing evaluation predominantly on maximal dataset examples, and paying less attention to the rest of the activation spectrum. Many polysemantic neurons appear monosemantic if you only look at top dataset examples, but are revealed to be polysemantic if you look at lower parts of the activation spectrum. To avoid this, we draw samples uniformly across the spectrum of feature activations,\nIn order to sample uniformly across the spectrum of feature activations, we divide the activation spectrum into 11 \"activation intervals\" evenly spaced between 0 activation and the maximum activation. We sample uniformly from these intervals.\nand score each interval separately in light of the overall hypothesis suggested by the feature.\nUnfortunately, this approach is labor intensive and so the number of scored samples is small. In total, 412 feature activation intervals were scored across 162 features and neurons.\nWe see that features are substantially more interpretable than neurons. Very subjectively, we found features to be quite interpretable if their rubric value was above 8. The median neuron scored 0 on our rubric, indicating that our annotator could not even form a hypothesis of what the neuron could represent! Whereas the median feature interval scored a 12, indicating that the annotator had a confident, specific, consistent hypothesis that made sense in terms of the logit output weights.\nAutomated Interpretability – Activations\nTo analyze features at a larger scale, we turned to automated interpretability\n. Following the approach of Bills\net al.\n, we have a large language model, Anthropic’s Claude, generate explanations of features using examples of tokens where they activate. Next, we have the model use that explanation to predict new activations on previously unseen tokens.\nIt's worth explicitly stating that our automated interpretability setup was designed to ensure that there's no leak of information about activation patterns, except for the explanation. For example, when predicting new activations, the model cannot see any true activations of that feature.\nLike with the human analysis, we used samples across the full range of activation intervals to evaluate monosemanticity.\nThis is distinct from the evaluation strategy of Bills\net al.\n, who calculated their correlation of predicted and true activations on a mixture of maximal dataset examples and random samples. For sparse features, which don't fire on most random samples, this effectively tests the model's ability to distinguish a feature's large activations from zero.\nConcretely, for each feature, we computed the Spearman correlation coefficient between the predicted activation and the true activations for 60 dataset examples made up of nine tokens each, resulting in 540 predictions per feature. While only using completely random sequences would be the most principled approach to scoring, half of the examples are from across the feature intervals to get a more accurate correlation. See\nthe appendix\nfor additional information including the use of importance scoring to precisely counter-weight the bias of providing tokens the feature fires for.\nIn agreement with the human analysis, Claude is able to explain and predict activations for features significantly better than for neurons.\nIn instances where Claude predicts a constant score, most often all 0s, a correlation can't be computed and we assign a score of zero which explains the uptick there.\nAutomated Interpretability – Logit Weights\nIn our earlier analysis of individual features, we found that looking at the logits is a powerful tool for cross-validating the interpretability of features. We can take this approach in automated interpretability as well. Using the explanations of features generated in the previous analysis, we ask a language model to predict if a previously unseen logit token is something the feature should predict as likely to come next. This is then scored against a 50/50 mix of top positive logit tokens and random other logit tokens. Randomly guessing would give a 50% accuracy, but the model instead achieves a 74% average across features, compared to a 58% average across neurons. Failures here refer to instances where Claude failed to reply in the correct format for scoring.\nActivation Interval Analysis\nIn addition to studying features as a whole, in our manual analysis we can zoom in on portions of the feature activation spectrum using the feature intervals. As before, a feature interval is the set of examples with activations closest to a specific evenly-spaced fraction of the max activation. So, rather than asking if a feature seems interpretable, we ask whether a range of activations is consistent with the overall hypothesis suggested by the full spectrum of the feature’s activation. This allows us to ask how interpretability changes with feature activation strength.\nHigher-activating feature intervals were more consistent with our interpretations than lower-activating ones. In particular:\nMany features show consistent activations across the entire activation spectrum.\nSome features show consistent activations across the top ~60% of the activation spectrum, and then quickly become less interpretable as we look to smaller and smaller activations.\nIt is possible that this is a sign that our features are not quite right. For instance, if one of our features is at a slight angle to the feature we’d really like to have learned, that can show up as inconsistent behavior in the lower activation intervals.\nCaveats\nOur manual and automated interpretability experiments have a few caveats:\nFeature activations are skewed towards the lower intervals.\nMost feature activations are quite small, and so fall in the lower (and less-interpretable) feature intervals. That said, as we found in our detailed analysis, most of the\neffect\nof a feature\nwith respect to the effect it has on the model outputs see e.g. the activation expected value plot in\nArabic Feature's Activations Specificity Analysis\n.\nis due to its higher activations, which are quite interpretable.\nThe evaluated features are sampled uniformly from among the set of all features, and interpretability might be correlated with importance.\nOne could imagine evaluating the features with the largest magnitude of activations, or the largest effect when ablated. Our sense is that these would be\nmore interpretable\nthan the randomly sampled features we considered. One could imagine a situation where the most important features in some sense were systematically less interpretable, though inspection of the\nmost active features\nsuggests the opposite.\nBased on our inspection of many features in the visualization, we believe these caveats do not affect the experimental results. We encourage interested readers to open the visualization for\nA/1\nand the\ncorresponding neurons\n. You can sort by ‘random’ to get an unbiased sample and do your own version of the above experiment, or sort features by importance metrics such as max activation and max density to evaluate the final caveat above.\nHow much of the model does our interpretation explain?\nWe now turn to the question we're least able to answer – to what extent do these seemingly interpretable features represent the \"full story\" of the MLP? One could imagine posing this question in a variety of ways. What fraction of the MLP loss contribution have we made interpretable? How much model behavior can we understand? If there really are some discrete set of \"true features\", what fraction have we discovered?\nOne way to partly get at this question is to ask how much of the loss is explained by our features. For\nA/1\n, the run we've focused most on in this paper, 79% of the log-likelihood loss reduction provided by the MLP layer is recovered by our features. That is, the additional loss incurred by replacing the MLP activations with the autoencoder's output is just 21% of the loss that would be incurred by zero ablating the MLP. This loss penalty can be reduced by using more features, or using a lower L1 coefficient. As an extreme example,\nA/\n5\n(\nn_learned_sparse=131,072\n,\nl1_coefficient=0.004\n) recovers 94.5% of log-likelihood loss.\nThese numbers should be taken with a significant grain of salt. The biggest issue is that framing this question in terms of fraction of loss may be misleading – we expect there to be a long-tail of features such that as the fraction of loss explained increases, more and more features are needed to explain the residual. Another issue is that we don't believe our features are completely monosemantic (some polysemanticity may be hiding in low activations), nor are all of them necessarily cleanly interpretable. With all of that said, our\nearlier analyses\nof individual features (e.g. the Arabic feature, base64 feature, etc.) do show that\nspecific\ninterpretable features are used by the model in interpretable ways – ablating them decreases probabilities in the appropriate way, and artificially activating them causes a corresponding behavior. This seems to confirm that the 79% of loss recovered is measuring something real, despite these caveats.\nIn principle, one could use automated interpretability to produce a better measure here: replacing activations with those predicted from explanations. (We believe others in the community have recently been considering this!) The naive versions of this would be quite computationally expensive,\nNaively, simulating a layer with 100k features would be 100,000 times more expensive than sampling a large language model such as Claude 2 or GPT-4 (we'd need to sample the explaining large model once for every feature at every token). At current prices, this would suggest simulating 100k features on a single 4096 token context would cost $12,500–$25,000, and one would presumably need to evaluate over many contexts.\nalthough there may be approximations. More generally, there is a much broader space of possibilities here. Perhaps there's a principled way to do this analysis in terms of single features – there are significant conceptual issues, but one might be able to formalize earlier notions of \"feature importance\" as an independent loss contribution a feature makes, and then analyze how much of that specific feature can be recovered.\nOverall, we view the problem of measuring the degree to which a feature-based interpretation explains a model to be an important open question, where significant work is necessary on both defining metrics and finding efficient ways to compute them.\nDo features tell us about the model or the data?\nA model's activations reflect two things: the distribution of the dataset and the way that distribution is transformed by the model. Dictionary learning on activations thus mixes data and model properties, and intriguing properties of learned features may be attributed to either or both sources. Correlations in the data can persist after application of the first part of the model (up to the MLP), and it is in theory possible that the intriguing features we see are merely artifacts of dataset correlations projected into a different space. However, the use of those features by the second half of the model (MLP downprojection and unembedding) are not an input to dictionary learning, so the interpretability of the downstream effects of those features must be a property of the model.\nTo assess the effect of dataset correlations on the interpretability of feature activations, we run dictionary learning on a version of our one-layer model with random weights.\nWe generate a model with random weights by randomly shuffling the entries of each weight matrix of the trained transformer used in Run A. This guarantees that the distributions of individual weights match, and differences are due to structure.\nThe resulting features are\nhere\n, and contain many single-token features (such as \"span\", \"file\", \".\", and \"nature\") and some other features firing on seemingly arbitrary subsets of different broadly recognizable contexts (such as\nLaTeX\nor\ncode\n). However, we are unable to construct interpretations for the non-single-token features that make much sense and invite the reader to examine feature visualizations from the model with randomized weights to confirm this for themselves. We conclude that the learning process for the model creates a richer structure in its activations than the distribution of tokens in the dataset alone.\nTo assess the interpretability of the downstream feature effects, we again use the three main approaches of the previous section:\nLogit weight inspection.\nThe logit weights represent the effect of each feature on the logits, and, as demonstrated in our earlier\ninvestigations of individual features\n, they are consistent with the feature activations for the base64, Arabic, and Hebrew features. The reader is invited to inspect logit weights for all features in the visualization.\nFeature ablation.\nWe set the value of a feature to zero throughout a context, and record how the loss on each token changes. Ablations are available for all features in the visualization.\nPinned feature sampling.\nWe artificially pinned the value of a feature to a fixed high number and then sample from the model. We find that the generated text matches the interpretation of the feature.\nThe empirical consistency of feature activations with their downstream effects across all these metrics provides evidence that the features found are being used by the model.\nPhenomenology\nUltimately, the goal of our work is to understand neural networks. Decomposition of models into features is simply a means to this end, and one might very reasonably wonder if it's genuinely advancing our overall goal. So, in this section, we'll turn attention to the lessons these features can teach us about neural networks. (We've taken to calling this work of leveraging our theoretical understanding to reason about model properties\nphenomenology\nby analogy to\nphenomenology in physics\n, and the\n2019 ICML workshop\non phenomena in deep learning.)\nOne way to do this would be to give a detailed discussion of the features we've found (similar to\n), but we believe the best way to get a sense of the features we discovered is simply to\nbrowse the interface\nfor exploring features which we've published alongside this paper. The features we find vary enormously, and no concise summary will capture their breadth. Instead, we will largely focus on more abstract properties and patterns that we notice. These abstract properties will be able to inform – although by no means answer – questions like \"Are these the\nreal\nfeatures?\" and \"What is actually going on in a one-layer model?\"\nWe begin by discussing some basic motifs and observations about features. We'll then discuss how the features we relate compare to features in other dictionary learning runs and in other models. This will suggest that features are\nuniversal\nand that dictionary learning can be understood as a process of\nfeature splitting\nthat reflects something deep about the geometry of superposition. Finally, we'll explore how features connect together into \"finite state automata\" as systems that implement more complex behaviors.\nFeature Motifs\nWhat kinds of features do we find in our model?\nOne strong theme is the prevalence of\ncontext features\n(e.g. DNA, base64) and\ntoken-in-context features\n(e.g.\nthe\nin mathematics –\nA/0/341\n,\n<\nin HTML –\nA/0/20\n).\nFrom a purely theoretical lens, attention heads can largely implement \"three point functions\" (with two inputs, and an output). MLP layers are well positioned to instead implement N-token conjunctions, perhaps the most extreme of which are context features or token-in-context features. Thus, it is perhaps natural that we see many of these.\nThese have been observed in prior work (\ncontext features\ne.g.\n;\ntoken-in-context features\ne.g.\n;\npreceding observations\n), but the sheer volume of token-in-context features has been striking to us. For example, in\nA/4\n, there are over a hundred features which primarily respond to the token \"the\" in different contexts.\nToken-in-context features may offer a significant opportunity for simplifying analysis of the model – as Elhage\net al.\nnote, it may be possible to understand these features as two-dimensional family of features parameterized by a context and a token.\nOften these features are connected by\nfeature splitting\n(discussed in the next section), presenting as pure context features or token features in dictionaries with few learned features, but then splitting into token-in-context features as more features are learned.\nAnother interesting pattern is the implementation of what seem to be \"trigram\" features, such as a feature that predicts the\n19\nin\nCOVID-19\n(\nA/2/12310\n). Such features could in principle be implemented with attention alone, but in practice the model uses the MLP layer as well. We also see features which seem to respond to specific, longer sequences of tokens. These are particularly striking because they may implement \"memorization\" like behavior – we'll discuss this more later.\nFinally, it's worth noting that all the features we find in a one-layer model can be interpreted as \"action features\" in addition to their role as \"input features\". For example, a base64 feature can be understood both as activating in response to base64 strings, and also as acting to increase the probability of base64 strings. The \"action\" view can clarify some of the token-in-context features: the feature\nA/0/341\npredicts noun phrases in mathematical text, upweighting nouns like\ndenominator\nand adjectives like\nlatter\n. Consequently, while it activates most strongly on\nthe\n, it also activates on adjectives like\nspecial\nand\nthis\nwhich are also followed by noun phrases. This dual interpretation of features can be explored by browsing our interface. Several papers have previously explored interpreting neurons as actions (e.g.\n), and one-layer models are particularly suited to this, since it's a particularly principled way to understand the last MLP layer, and the only MLP in a one-layer model is the last layer.\nFeature Splitting\nOne striking thing about the features we’ve found is that they appear in\nclusters\n. For instance, we observed above multiple base64 features, multiple Arabic script features, and so on. We see more of these features as we increase the total number of learned sparse features, a phenomenon we refer to as\nfeature splitting\n. As we go from 512 features in A/0 to 4,096 features in A/1 and to 16,384 features in A/2, the number of features specific to base64 contexts goes from 1 to 3 to many more.\nTo understand how the geometry of the dictionary elements correspond to these qualitative clusters, we do a 2-D UMAP on the combined set of feature directions from A/0, A/1, and A/2.\nWe see clusters corresponding to the base64 and Arabic script features, together with many other tight clusters from specific contexts and a variety of other interesting geometric structures for other features. This confirms that the qualitative clusters are reflected in the geometry of the dictionary: similar features have small angles between their dictionary vectors.\nWe conjecture that there is some idealized set of features that dictionary learning would return if we provided it with an unlimited dictionary size. Often, these \"true features\" are clustered into sets of similar features, which the model puts in very tight superposition. Because the number of features is restricted, dictionary learning instead returns features which cover approximately the same territory as the idealized features, at the cost of being somewhat less specific.\nIn this picture, the reason the dictionary vectors of conceptually similar features are similar is that they are likely to produce similar behaviors in the model, and so should be responsible for similar effects in the neuron activations. For instance, it would be natural for a feature that fires on periods to predict tokens with a leading space followed by a capital letter. If there are multiple features that fire on periods, perhaps on periods in somewhat different contexts, these might all predict tokens with a leading space, and those predictions might well involve producing similar neuron activations. The combination of features being highly correlated and having similar \"output actions\", causes real models to have both denser and more structured superposition than what we observed in our previous toy models work\n.\nToy Models\nconsidered correlated features (see in particular\norganization of correlated features\nand\ncollapsing of correlated features\n), but only features which were correlated in\nwhether they were active\n(and not their value if active), and had nothing analogous to the similar \"output actions\" described here. Nonetheless,\nToy Models'\nexperiments may be a useful intuition pump, especially in noticing the distinction between similar features in superposition vs features collapsing into a single broader feature.\nIf this picture is true, it would be important for a number of reasons. It suggests that determining the \"correct number of features\" for dictionary learning is less important than it might initially seem. It also suggests that dictionary learning with fewer features can provide a \"summary\" of model features, which might be very important in studying large models. Additionally, it would explain some of the stranger features we observe in the process of dictionary learning, suggesting that these are either \"collapsed\" features which would make sense if split further (see\n\"Bug\" 1: Single Token Features\n), or else highly-specific \"split\" features which do in fact make sense if analyzed closely (see\n\"Bug\" 2: Multiple Features for a Single Context\n). Finally, it suggests that our basic theory of superposition in toy models is missing an important dimension of the problem by not adequately studying highly correlated and \"action sharing\" features.\nExample: Mathematics and Physics Features\nIn this example, our coarsest run (with 512 learned sparse features) has three features describing tokens in different technical settings. Using the\nmasked cosine similarity\nFor each pair of features, we compute the cosine similarity between their activations on the subset of tokens for which one of them fires. We repeat this, restricting to the subset on which the other fires, and take the greater of the two. We draw a connection between the features if this measure exceeds ~0.4. This threshold was chosen to balance producing a small enough graph to visualize while also showing some of the richness of feature splitting. We omit the ultralow density features from this analysis.\nbetween feature activations, we are able to identify how these features refine and split in runs with more learned sparse features.\nWhat we see is that the finer runs reveal more fine-grained distinctions between e.g. concepts in technical writing, and distinguish between the articles\nthe\nand\na\n, which are followed by slightly different sets of noun phrases. We also see that the structure of this refinement is more complex than a tree: rather, the features we find at one level may both split and merge to form refined features at the next. In general though, we see that runs with more learned sparse features tend to be more specific than those with fewer.\nIt's worth noting that these more precise features reflect differences in model predictions as well as activations. The general\nthe\nin mathematical prose feature (\nA/0/341\n) has highly generic mathematical tokens for its top positive logits (e.g. supporting\nthe denominator\n,\nthe remainder\n,\nthe theorem\n), whereas the more finely split machine learning version (\nA/2/15021\n) has much more specific topical predictions (e.g.\nthe dataset\n,\nthe classifier\n). Likewise, our abstract algebra and topology feature (\nA/2/4878\n) supports\nthe quotient\nand\nthe subgroup\n, and the gravitation and field theory feature (\nA/2/2609\n) supports\nthe gauge\n,\nthe Lagrangian\n, and\nthe spacetime\n.\nFeatures which seemed like Bugs\n\"Bug\" 1: Single-Token Features\nWhen we limit dictionary learning to use very few learned sparse features, the features that emerge sometimes look quite strange. In particular, there are a large number of high-activation magnitude features which each only fire on a single token, and which seem to fire on every instance of that token. Such features are strange because the model could achieve the same effect entirely by learning different bigram statistics, and so should have no reason to devote MLP capacity to these. Similar features were also recently observed in a report by Smith\n.\nWe believe that feature splitting explains this phenomenon: the model hasn’t learned a single feature firing on the letter\nP\n,\nThe features fire on the token \" P\" of words where it appears as the first token, such as\n[ P][attern]\n.\nfor instance. Rather, it’s learned many features which fire on\nP\nin different contexts\nIn this instance we found the refined\nP\nfeatures by manual inspection rather than by cosine similarity.\n, with correspondingly different effects on the neuron activations and output logits (see below). At a sufficiently coarse level dictionary learning cannot tell the difference between these, but when we allow it to use more learned sparse features, the features split and refine into a zoo of different\nP\nfeatures that fire in different contexts.\n\"Bug\" 2: Multiple Features for a Single Context\nWe also observed the converse, where multiple features seemed to cover roughly the same concept or context. For example, there were three features in\nA/1\nwhich fired on (subsets of) base64 strings, and predicted plausible base64 tokens like\nzf\n,\nmF\n, and\nGp\n. One of these features was discussed in detail\nearlier\n, where we showed it fired for base64 strings. But we also observed that it didn't fire for all base64 strings – why? And what are the other two features doing? Why are there three?\nIn\nA/0\n(with 512 features), the story is simple. There is only one base64-related feature,\nA/0/45\n, which seems to activate on\nall\ntokens of base64-encoded strings. But in\nA/1\n, that feature\nsplits\ninto three different features whose activations seem to jointly cover those of\nA/0/45\n:\nTwo of these features seem relatively straightforward.\nA/1/2357\nseems to fire preferentially on letters in base64, while\nA/1/2364\nseems to fire preferentially on digits.\nComparing the logit weights of these features reveals that they predict largely the same sets of tokens, with one significant difference: the feature that fires on digits has much lower logit weights for predicting digits. Put another way, if the present token is made of digits, the model will predict that the next token is a non-digit base64 token.\nWe believe this is likely an artifact of tokenization! If a single digit were followed by another digit, they would have been tokenized together as a single token;\n[Bq][8][9][mp]\nwould never occur, as it would be tokenized instead as\n[Bq][89][mp]\n. Thus even in a random base64 string, the fact that the current token is a single digit gives information about the next token.\nBut what about the third feature,\nA/1/1544\n? At first glance, there isn't an obvious rule for when it fires. But if we look more closely, we notice that it seems to respond to base64 strings\nwhich encode ASCII text\n.\nOur initial clue that\nA/1/1544\nmight fire on base64 strings encoding ASCII text was the token\nICAgICAg\nwhich this feature particularly responds to, and corresponds to six spaces in a row.\nIf we look at the top dataset examples for each feature, we find that examples for\nA/1/1544\ncontain substrings which decode as ASCII, while none of the top activating examples for\nA/1/2357\nor\nA/1/2364\ndo:\nDetermining when a dataset example encodes ASCII text is somewhat subtle because base64 can only be decoded to ASCII in groups of four characters, since four base64 characters\nencode triples\nof ASCII characters. Thus, we select substrings which – when decoded with the python base64 library – contain the maximal number of printable ASCII characters.\nThis pattern of investigation, where one looks at coarser sets of features to understand categories of model behavior, and then at more refined sets of features to investigate the subtleties of that behavior, may prove well adapted to larger models where the feature set is expected to be quite large.\nIt's also worth noting how dictionary learning features were able to\nsurprise us\nhere. Many approaches to interpretability are top-down, and look for things we expect. But who would have known that models not only have a base64 feature, but that they distinguish between distinct kinds of base64 strings? This reminds us of cases like high-low frequency detectors\nor multimodal neurons\nwhere surprising and unexpected features were discovered in vision models.\nUniversality\nOne of the biggest \"meta questions\" about features is whether they're\nuniversal\n– do the same features form across different models? This question is generally important because it bears on whether the hard-earned lessons from studying one model will generalize to others. But it's especially important in the context of attempting to extract features from superposition because universality could provide significant evidence that the features we're extracting are \"real\", or at least reproducible.\nIn what sense does universality suggest features are \"real\"? One basic observation is that they suggest the features we're finding are not just artifacts of the dictionary learning process – or at least that if they come from the dictionary learning process, it's in some consistent way. But there are also several deeper ways in which it's suggestive. It means that, whatever the source of the features, we can talk about features as replicable, reliable, recurring units of analysis. It's also just a surprising observation that one would expect if a strong version of the features in superposition hypothesis was true and models were literally representing some finite, discrete set of features in superposition.\nEarlier, we saw that all the features we performed detailed analyses of (e.g. the Arabic feature, or base64 feature) were universal between two one-layer models. But is this true for typical features in our model? And how broadly is it true – do we only observe the same feature if we train models of the same architectures on the same dataset, or do these features also occur in more divergent models? This section will seek to address these two questions. The first subsection will quantitatively analyze how widespread universality is between the two one-layer models we studied, while the second will compare the features we find to others reported in the literature in search of a stronger form of universality.\nWe observe substantial universality of both types.\nIn fact, we found some features so universal that we began to take it for granted as a basic tool in our workflow of evaluating dictionary learning runs. For example, the base64 feature – which we previously observed in SoLU models – was so consistently universal that its presence was a useful debugging heuristic.\nAt a high-level, this makes sense: if a feature is useful to one model in representing the dataset, it's likely useful to others, and if two models represent the same feature then a good dictionary learning algorithm should find it.\nComparing features between two one-layer transformers\nTo compare features from different models, we need model-independent ways to represent a feature.\nOne natural approach is to think of a feature as a function assigning values to datapoints; two features would be similar in this sense if they take similar values over a diverse set of data. This general approach has been explored by a number of prior papers (e.g.\n). In practice, this can be approximated by representing the feature as a vector, with indices corresponding to a fixed set of data points. We call the correlations between these vectors the\nactivation similarity\nbetween features.\nA second natural approach is to think of a feature in terms of its downstream effects; two features would be similar in this sense if their activation changes their models' predictions in similar ways. In our one-layer model, a simple approximation to this is the logit weights. This approximation represents each feature as a vector with indices corresponding to vocabulary tokens. We call the correlations between these vectors the\nlogit weight similarity\nbetween features.\nThese two notions of similarity correspond to the correlations of the points in the two scatter plots we used when\nanalyzing individual features\nearlier. We've reproduced the plots for the Arabic feature below:\nFor each feature in run\nA/1\n, we find the closest feature by activation similarity in run\nB/1\n, which is a different dictionary learning run trained on different activations from a different transformer with different random seeds but otherwise identical hyperparameters. We find that many features are highly similar between models, with features in\nA/1\nhaving a median activation correlation of 0.72 with the most similar feature from\nB/1\n. (We perform the same analysis finding the closest neurons between the transformers, and find significantly less similarity, with median activation correlation 0.46.) The features with low activation correlation between models may represent different \"feature splittings\" in the dictionaries learned or different \"true features\" learned by the base models.\nA natural next question is whether features that fire on the same tokens also have the same logit effects. That is, how well do activation similarity and logit weight similarity agree?\nSome gap between the two is visible for the Arabic feature above: the \"important tokens\" for the features' effects (the ones in Arabic script) are upweighted by features from both models, but there is a large cloud of tokens with smaller effects that appear to almost be isotropic noise, resulting in a logit weight correlation of just 0.23, significantly below the activation correlation of 0.91.\nIn the scatterplot below, we find that this kind of disagreement is widespread.\nThe most dramatic example of this disparity is for the features\nA/1/3949\nand\nB/1/3321\n, with an activation correlation of 0.98 but a\nnegative\nlogit weight correlation. These features fire on\npone\n(and occasionally on\npgen\nand\npcbi\n) as abbreviations for the journal name PLOSOne in citations, like\n@\npone\n.0082392\n, and predict the\n.\nthat follows.\nThe feature is bimodal, monosemantic in the larger of the modes, and fires on 0.02% of tokens. This means that at least 1 in 10,000 tokens in the Pile dataset are abbreviations for PLoS journals in citations! This is an example of how inspecting features can reveal properties of the dataset, in this case the strong bias of the Pile towards scientific content.\nZooming in on the logit weight scatterplot (inset in the figure above), we see that only the\n.\ntoken has high logit weight in both models, and that every other token is in the 'interference' portion of the logit weight distribution. Indeed, the model may simply not care about what the feature does to tokens which were already implausible because they are suppressed by the direct path, attention layer, or other features of the MLP.\nWe want to measure something more like \"the actual effect a feature has on token probabilities.\" One way to get at this would be to compute a vector of ablation effects for every feature on every data point; pairs of features whose ablations hurt the model's predictions on the same tokens must have been predicting the same thing. Unfortunately, this would be rather expensive computationally. Instead, we scale the activation vector of a feature by the logit weights of the tokens that empirically come next in the dataset to produce an\nattribution vector\n.\nSuppose feature\nf_i\nhas logit weights\nv_{ik}\nfor\nk \\in \\{1,\\ldots,n_{\\text{vocab}}\\}\n. At a given token\nt_j\n, we compute the activation of the feature\nf_i(t_j)\nand multiply it by the logit weight\nv_{it_{j+1}}\nof the token\nt_{j+1}\nthat comes next to get an attribution score of\nf_i(t_j)v_{it_{j+1}}\n. The attribution vector is given by stacking the attribution scores for a random sampling of datapoints. This approximates the classic attribution method of multiplying the gradient by the activation, differing in that we ignore the denominators of the softmax and the layer norm.\nCorrelations between those vectors provide an\nattribution similarity\nthat combines both the activity of the feature with the effect it has on the loss. We find that the attribution similarity correlates quite highly with the activation similarity, meaning that features that were coactive between models were useful at predicting the same tokens.\nIn light of this, we feel that the activation correlation used throughout the paper is in fact a good proxy for both notions of universality in the context of our one-layer models.\nComparing features with the literature\nSo far, we've established that many of our features are universal in a limited sense. Features found in one of our transformers can also be found in an alternative version trained with a different random seed. But this second model has an identical architecture and was trained on identical data. This is the most minimal version of universality one could hope for. Despite this, we believe that many of the features we've found are universal in a deeper sense, because very similar features have been reported in the literature before.\nThe first comparison which struck us is that many features seem quite similar to neurons we previously found in one-layer SoLU models, which use an activation function designed to make neurons more monosemantic\n. In particular, we observed a base64 neuron, hexadecimal neuron, and all caps neuron in our SoLU investigations, and base64 (\nA/0/45\n), hexadecimal (\nA/0/119\n), and all caps (\nA/0/317\n) features here. Discussion of some of these neurons can be found in\nSection 6.3.1\nof the SoLU paper.\nWe also find many features similar to Smith\n, who applies dictionary learning to the residual stream. In addition to us also observing preponderance of single token features they note (see\nour interpretation\nof this phenomenon), we find a similar German detector (e.g.\nA/0/493\n) and similar title case detectors (e.g.\nA/0/508\n). Likewise, we find a number of features similar to Gurnee\net al\n.\n, including a \"prime factors\" feature  (\nA/4/22414\n) and a French feature (\nA/0/14\n).\nAt a more abstract level, many features we find seem similar to features reported in multimodal models by Goh\net al.\n. For example, we find many similar features including an Australia feature (\nA/3/16085\n), Canada feature (\nA/3/13683\n), Africa feature (\nA/3/14490\n), and Israel-Palestine feature (\nA/3/739\n) which predict locations in those regions when grammatically appropriate. This vaguely mirrors \"region neurons\" reported by Goh\net al.\n's paper. For other families of features, the parallels are less clear. For example, one of the most striking results of Goh\net al.\nwas person detector neurons (similar to famous results in neuroscience). We find some features that are person detectors in very narrow contexts, such as responding to a person’s name and predicting appropriate next words, or predicting their name (e.g.\nA/1/3240\nis somewhat similar to Goh\net al.\n's Trump neuron), but they seem quite narrow. We also don't find features that seem clearly analogous to Goh\net al.\n's emotion neurons.\n\"Finite State Automata\"\nOne of the most striking phenomena we've observed in our study of the features in one-layer models is the existence of \"finite state automata\"-like assemblies of features. These assemblies aren't circuits in the conventional sense – they're formed by one feature increasing the probability of tokens, which in turn cause another feature to fire on the next step, and so on.\nThe \"finite state automata\"-esque feature assemblies are also different from circuits in that the model didn't learn them to work together. Rather, in the course of learning to autoregressively model text, it learned features that interact via the token stream because of patterns in the real datasets. In contrast, a language model trained with reinforcement learning might have systems like this – circuits whose feature components interact via generated tokens – which co-evolved and adapted to work together during RL training.\nThe simplest example of this is features which excite themselves on the next token, forming a single node loop. For example, a base64 feature increases the probability of tokens like\nQg\nand\nzA\n– plausible continuations which would continue to activate it.\nIt's worth noting that these examples are from\nA/0\n, a dictionary learning run which is not overcomplete (the dictionary dimensionality is 512, equal to the transformer MLP dimension). As we move to runs with larger numbers of features, the central feature will experience feature splitting, and become a more complex system.\nLet's now consider a two-node system for producing variables in \"all caps snake case\" (e.g.\nARRAY_MAX_VALUE\n). One node (\nA/0/207\n) activates on the all caps text tokens, the other (\nA/0/358\n) on underscores:\nThis type of two-node system is quite common for languages where Unicode characters are sometimes split into two tokens. (Again, with more feature splitting, these would expand into more complex systems.)\nFor example, Tamil Unicode characters (\nblock U+0B80–U+0BFF\n) are typically split into two tokens. For example, the character \"ண\" (\nU+0BA3\n) is tokenized as\n\\xe0\\xae\nfollowed by\n\\xa3\n.  The first part (\n\\xe0\\xae\nor\n\\xe0\\xaf\n) roughly specifies the Unicode block, while the second component specifies the character within that block. Thus, it's natural for the model to alternate between two features, one for the Unicode prefix token, and one for the suffix token.\nA more complex example is Chinese. While many common Chinese characters get dedicated tokens, many others are split. This is further complicated by Chinese characters being spread over many Unicode blocks, and those blocks being large and cutting across many logical blocks specified in terms of bytes. To understand the state machine the model implements to handle this, the key observation is that complete characters are similar to the \"suffix\" part of a split character: both can be followed by either a new complete character, or a new prefix. Thus, we observe two features, one of which fires on\neither\ncomplete characters or the suffix (predicting either a new complete character, or a prefix), while the other only fires on the prefixes and predicts suffixes.\nLet's now consider a very simple four node system which models HTML. The \"main path\" through it is:\nA/0/20\nfires on open tags and predicts tag names\nA/0/0\nfires on tag names and predicts tag closes\nA/0/30\nfires on tag closes and predicts whitespace\nA/0/494\nfires on whitespace and predicts new tag opens.\nA prototypical sample this might generate is something like\n<div>\\n\\t\\t<span>\n.\nThe full system can be seen below:\nKeep in mind that we're focusing on the\nA/0\nfeatures where this is very simple – if we looked at\nA/1\n, we'd find something much more complex! One particularly striking shortcoming of the\nA/0\nfeatures is that they don't describe what happens when\nA/0/0\nemits a token like\nhref\n, which leads to a more complex state.\nIt's important to note that these features can be quite contextual. There are several features related to IRC transcripts which form a totally different finite state automata like system:\nA prototypical sample this might generate is something like\n<nickonia_> lol ubuntu ;)\n. Presumably the Pile dataset heavily represents IRC transcripts about linux.\nOne particularly interesting behavior is the apparent memorization of specific phrases. This can be observed only in runs with relatively large numbers of features (like\nA/4\n). In the following example, a sequence of features seem to functionally memorize the bolded part of the phrase\nMERCHANTABILITY or FITNESS FOR A\nPARTICULAR PURPOSE\n. This is a relatively standard legal language, and notably occurs in the file headers for popular open source software licenses, meaning the model likely saw it many times during training.\nThis seems like an example of the mechanistic theory of memorization we described in Henighan\net al.\n– we observe features which appear to be relatively binary and respond to a very specific situation. This might also be seen as an instance of mechanistic anomaly detection\n: the model behaves differently in a specific, narrow case. It's somewhat surprising that something so narrow can be found in a model with only 512 neurons; from this perspective it's an interesting example of superposition's ability to embed many things in few neurons. On the other hand, because these mechanisms are buried deep in superposition, they are likely very noisy.\nRelated Work\nSuperposition and attempts to resolve it have deep connections to many lines of research, including general investigations of interpretable features, linear probing, compressed sensing, dictionary learning and sparse coding, theories of neural coding, distributed representations, mathematical frames, vector symbolic architectures, and much more. Rather than attempt to do justice to all these connections here, we refer readers to the\nrelated work section\nof Toy Models of Superposition\nwhere we discuss these topics in depth, and also to our essay\nDistributed Representations: Composition & Superposition\n. Instead, we'll focus our discussion on work connected to our attempts to\nsolve\nsuperposition, and also more recent advancements in our understanding of superposition.\nSuperposition\nSince we published Toy Models of Superposition there has been significant further work attempting to better understand superposition. We briefly summarize below.\nIs superposition real?\nGurnee\net al.\ndemonstrate some compelling examples of features which may be in superposition, using sparse linear probes. Separately, an exchange between Li\net al.\nand Nanda\net al.\nseems like an update on whether the general picture of features as directions is correct; Li\net al.\nseemed to show that it wasn't, putting the hypothesis in jeopardy, which was then resolved by Nanda\net al\n.\nWhen and why does superposition occur?\nScherlis\net al.\nprovide a mathematical framework for thinking about monosemanticity vs polysemanticity. lsgos\nexplored the effects of dropout on toy models of superposition.\nMemorization\n– In Henighan\net al.\n, we studied the same toy model as in Toy Models of Superposition, but this time trained on many repetitions of finite-sized datasets. We found that small datasets are\nmemorized\nin superposition, instead of generalizing features in the case of large datasets. Hobbhahn\nreplicated some of these findings, and further showed extensions to other settings including bottlenecks between layers (analogous to the residual stream between MLP layers). Subsequently, in a monthly update we\nexamined\nthe boundary between the memorization and generalization regimes and found a sharp phase transition, as well as dataset clustering in superposition on the memorization side of the boundary.\nDisentanglement and Architectural Approaches\nThere is also a rich and related literature on disentanglement, which seeks to find representations of data that separate out (disentangle) conceptually-distinct phenomena influencing the data. In contrast to superposition, this work typically seeks to find a number of factors of variation or features which are\nequal\nto the dimensionality of the space being represented, whereas superposition seeks to find more.\nThis is often approached as an architecture/training-time problem. For instance, Kim & Mnih\nproposed a method that pushes variational autoencoders to disentangle factors by encouraging independence across dimensions. Similarly, Chen\net al.\ndeveloped a Generative Adversarial Network approach that attempts to disentangle factors by maximizing the mutual information between a small subset of factors and the dataset. And Makhzani & Frey\nuse a TopK activation to encourage sparsity and hence disentanglement. See Bengio\net al.\nand Räuker\net al.\nfor a discussion of other such approaches.\nFramed this way, some architectural approaches to superposition may also be understood as attempts at disentanglement. For instance, in Elhage\net al.\n, we proposed the SoLU activation function, which increases the number of interpretable neurons in transformers by encouraging features to align to the neuron basis. Unfortunately, it appears that training models with the SoLU activation function may make some neurons more interpretable at the cost of making others even less interpretable than before.\nSimilarly, Jermyn\net al.\nstudied MLP layers trained on a compressed sensing task and found multiple equal-loss minima, with some strongly polysemantic and others strongly monosemantic. This suggested that training interventions could steer models towards more monosemantic minima, though subsequent investigations on more realistic tasks suggested that the equal-loss property was specific to the chosen task.\nThese two examples of attempts to tackle superposition through architecture, and the challenges they encountered, highlight a key distinction between the problems of disentanglement and that of superposition: disentanglement fundamentally seeks to ensure that the dimensions in the model’s latent space are disentangled, whereas superposition hypothesizes that this disentanglement typically hurts performance (since success would require throwing away many features), and that models will typically respond to disentangling interventions by making some features more strongly entangled (as was found by both Mahinpei\net al.\nand Elhage et al. 2022\nin somewhat different contexts).\nDictionary Learning and Features\nOur work builds on a longer tradition of using dictionary learning and sparse autoencoders to decompose neural network activations.\nEarly work in this space focused on word embeddings and other non-transformer neural networks. Faruqui\net al.\nand\nArora\net al.\nboth found linear structure in word embeddings using sparse coding approaches. Subramanian\net al.\nsimilarly found linear factors for word embeddings, in this case using a sparse autoencoder. Zhang\net al.\nsolved a similar problem using methods from dictionary learning while Panigrahi\net al.\napproached this with Latent Dirichlet Allocation.\nMore recently, a number of works have applied dictionary learning methods to transformer models. Yun\net al.\napplied dictionary learning to the residual stream of a 12-layer transformer to find an undercomplete basis of features.\nAt this point, our work in\nToy Models\nadvocated for dictionary learning as a potential approach to superposition. This motivated a parallel investigation by our colleagues Cunningham\net al.\n, published as a series of interim reports\nwith very similar themes to this paper, culminating in a manuscript\n. We've been excited to see so many corroborating findings between our work.\nIn their interim reports, Sharkey\net al.\nused sparse autoencoders to perform dictionary learning on a one-layer transformer, identifying a large (overcomplete) basis of features. (Sharkey\net al.\ndeserve credit for focusing on dictionary learning and especially the sparse autoencoder approach, while our investigation was only exploring it as one of several approaches in parallel.) This work was then partially replicated by Cunningham & Smith\nand Huben\n. Next, Smith\nused an autoencoder to find features in one MLP layer of a six-layer model. The resulting features appear interpretable, e.g. detecting ‘$’ in the context of LaTeX equations. In follow up work, Smith then extended this approach to the residual stream of the same model, identifying a number of interesting features (see\nearlier discussion\n). Building on these results, Cunningham\napplied autointerpretability techniques from Bills\net al.\nto features in the residual stream and an MLP layer of the same six-layer model, finding that the features discovered by the sparse autoencoder are substantially more interpretable than neurons.\nDiscussion\nTheories of Superposition\nComing into this work, our understanding of superposition was mostly informed by Toy Models\n. This gave us a picture one might call the\nisotropic superposition\nmodel. Features are discrete, one-dimensional objects which repel from each other due to interference, creating a roughly evenly spaced organization of feature directions.\nThis work has persuaded us that our previous model was missing something crucial. At a minimum, features seem to clump together in higher density groups of related features. One explanation for this (considered briefly by Toy Models) is that the features may have correlated activations – firing together. Another – which we suspect to be more central – is that the features produce\nsimilar actions\n. The feature which fires on single digits in base64 predicts approximately the same set of tokens as the feature firing on other characters in base64, with the exception of other digits; these similar downstream effects manifest as geometrically close feature directions.\nMoreover, it isn't clear that features need to be one-dimensional objects (encoding only some intensity). In principle, it seems possible to have higher-dimensional \"feature manifolds\" (see earlier discussion\nhere\n).\nThese hypotheses are not mutually exclusive. The convex hull of several correlated features might be understood as a feature manifold. On the other hand, some manifolds would not admit a unique description in terms of a finite number of one-dimensional features. (Perhaps this accounts for the continued\nfeature splitting\nobserved above.)\nNevertheless, these experiments have left us more confident that some version of the superposition hypothesis (and the linear representation hypothesis) is true. The number of interpretable features found, the way activation level seems to correspond to \"intensity\" or \"confidence,\" the fact that logit weights mostly make sense, and the observation of \"interference weights\": all of these observations are what you would expect from superposition.\nFinally, we note that in some of these expanded theories of superposition, finding the \"correct number of features\" may not be well-posed. In others, there is a true number of features, but getting it exactly right is less essential because we \"fail gracefully\", observing the \"true features\" at resolutions of different granularity as we increase the number of learned features in the autoencoder.\nAre \"Token in Context\" Features Real?\nOne of the most common motifs we found were \"token-in-context\" features. They also represent many of the features that emerge via feature splitting with increasing dictionary size. Some of these are intuitive – borrowing an example from\n, it makes sense to represent \"die\" in German (where it's the definite article) as distinct from \"die\" in English (where it means \"death\" or \"dice\").\nBut why do we see hundreds of different features for \"the\" (such as \"the\" in Physics, as distinct from \"the\" in mathematics)? We also observe this for other common words (e.g. \"a\", \"of\"), and for punctuation like periods. These features are not what we expected to find when we set out to investigate one-layer models!\nTo make the question a bit more precise, it is helpful to borrow the language and examples of local vs compositional representations\n. Individually representing token-context pairs (such as \"the\" in Physics) is technically a\n\"local code\"\n. The more intuitive way to represent this would instead be a\n\"compositional code\"\n– representing\n\"the\" as an independent feature from Physics. So the thing we really want to ask is why we're observing a local code, and whether it's really what's going on. There are two hypotheses:\nThe underlying transformer uses a compositional code, and a quirk of our dictionary learning scheme produces features using a local code.\nThe underlying transformer is genuinely using a local code (at least in part), and dictionary learning is correctly representing this.\nIf the former holds, then better dictionary learning schemes may help uncover a more compositional set of features from the same transformer. Local codes are sparser than compositional codes, and our L1 penalty may be pushing the model too far towards sparsity.\nHowever, we believe the second hypothesis is likely to hold to some extent. Let's consider the example of \"the\" in Physics again, which predicts noun phrases in Physics: if the model represented \"the\" and Physics context independently, it would be forced to have logits be the sum of \"upweight tokens which come after the\" and \"upweight tokens which occur in Physics\". But the model might wish to have \"sharper\" predictions than this, which is only possible with a local code.\nFuture Work\nScaling Sparse Autoencoders.\nScaling the application of sparse autoencoders to frontier models strikes us as one of the most important questions going forward. We're quite hopeful that these or similar methods will work – Cunningham\net al.\n's work\nseems to suggest this approach can work on somewhat larger models, and we have preliminary results that point in the same direction. However, there are significant computational challenges to be overcome. Consider an autoencoder with a 100× expansion factor applied to the activations of a single MLP layer of width 10,000: it would have ~20 billion parameters. Additionally, many of these features are likely quite rare, potentially requiring the autoencoder to be trained on a substantial fraction of the large model's training corpus. So it seems plausible that training the autoencoder could become very expensive, potentially even more expensive than the original model. We remain optimistic, however, and there is a silver lining –\nit increasingly seems like a large chunk of the mechanistic interpretability agenda will now turn on succeeding at a difficult engineering and scaling problem, which frontier AI labs have significant expertise in.\nScaling Laws for Dictionary Learning.\nIt's worth noting that there's enormous uncertainty about the dynamics of scaling dictionary learning and sparse autoencoders discussed above. As we make the subject model bigger, how does the ideal expansion factor change? (Does it stay constant?) How does the necessary amount of data change? The resolution of these questions will determine whether it's possible for this approach, if executed well, to scale up to frontier models. Ideally, we'd like to have\nscaling laws\nwhich could answer this.\nHow Can We Recognize Good Features?\nOne of the greatest challenges of this work is that we're \"wandering in the dark\" to some extent. We don't have a great, systematic way to know if we're successfully extracting high quality features.\nAutomated interpretability\nseems like a strong contender for solving this question. Alternatively, one might hope for some purely abstract definition (e.g. the\ninformation-based metric\nproposal), but we have not yet seen compelling signs of life for this on real data. It would also be helpful to have metrics beyond MMCS, activation similarity, and attribution similarity for comparing sets of features for the purposes of assessing consistency and universality.\nScalability of Analysis.\nSuppose that sparse autoencoders fully solve superposition. Do we have a home run to fully mechanistically understanding models? It seems clear that there would be at least one other fundamental barrier:\nscaling analysis\nof models, so that we can turn microscopic insights into a more macroscopic understanding. Again, one approach here could be automated interpretability. But delegating the understanding of AI to AI may not be fully satisfying, for various reasons. It is possible that there may be other paths based on discovering larger scale structure (see discussion\nhere\n).\nAlgorithmic Improvements for Sparse Autoencoders.\nNew algorithms refining the sparse autoencoder approach could be useful. One might explore the use of variational autoencoders (e.g.,\n), or sparsity promoting priors regularization techniques beyond a simple L1 penalty on activations (e.g.,\n), for example encouraging sparsity in the interactions between learned features in different layers. Earlier research has shown that noise injection can also increase neuron interpretability separately from an L1 penalty\n.\nAttentional Superposition?\nMany of the motivations for the presence of superposition in MLP layers\napply to self-attention layers as well. It seems conceivable that similar methods may extract useful structure from attention layers, although a clear example has not yet been established (e.g., see our\nMay\nand\nJuly\nUpdates). If this is true, addressing this may become a future bottleneck for the mechanistic interpretability agenda.\nTheory of Superposition and Features.\nMany fundamental questions remain for our understanding of superposition, even if the hypothesis is right in some very broad sense. For example, as discussed above, this work suggests extensions of the superposition hypothesis covering clusters of features with similar effects, or continuous families of features. We believe there is important work to be done in exploring the theory of superposition further, perhaps through the use of toy models.\nComments & Replications\nInspired by the original\nCircuits Thread\nand\nDistill's Discussion Article experiment\n, the authors invited several external researchers who we had previously discussed our preliminary results with to comment on this work. Their comments are included below.\nReplication & Tutorial\nNeel Nanda\nis an external mechanistic interpretability researcher. This is a summary of\na blog post\nreplicating and extending this paper, with\nan accompanying tutorial\nto load some trained autoencoders and practice interpreting a feature.\nThe core results of this paper seem to replicate. I trained a sparse autoencoder on the MLP layer of an open source 1 layer GELU language model, and a significant fraction of the latent space features were interpretable.\nI've open sourced two trained autoencoders and\na tutorial for how to use them, and how to interpret a feature\n. I've also open sourced a (very!) rough\ntraining codebase\n, along with some\nof the implementation details\nthat came up, and tips for training your own.\nI investigated\nhow sparse the decoder weights are in the neuron basis\nand find that they’re highly distributed, with 4% well explained by a single neuron, 4% well explained by 2 to 10, and the remaining 92% dense. I find this pretty surprising! Despite this, kurtosis shows the neuron basis is still privileged.\nI also exhibit\nsome case studies\nof the features I found, like a title case feature, and an \"and I\" feature\nI didn’t find any dead features, but more than half of the features form\nan ultra-low frequency cluster\n(frequency less than 1e-4). Surprisingly, I find that this cluster is almost all the same feature (in terms of encoder weights, but not in terms of decoder weights). On one input 95% of these ultra rare features fired!\nThe same direction forms across random seeds, suggesting it's a true thing about the model and not just an autoencoder artifact\nI failed to interpret what this shared direction was\nI tried to fix the problem by training an autoencoder to be orthogonal to this direction but it still forms an ultra-low frequency cluster (which all cluster in a new direction)\nOne question from this work is whether the encoder and decoder should be tied. I find that, empirically,\nthe decoder and encoder weights for each feature are moderately different\n, with median cosine similarity of only 0.5, which is empirical evidence they're doing different things and should not be tied. Conceptually, the encoder and decoder are doing different things: the encoder is detecting, finding the optimal direction to project onto to detect the feature, minimising interference with other similar features, while the decoder is trying to represent the feature, and tries to approximate the “true” feature direction regardless of any interference.\nAuthor Contributions Statement\nInfrastructure, Tooling, and Core Algorithmic Work\nGeneral Infrastructure\n– The basic framework for our dictionary learning work was built and maintained by Adly Templeton, Trenton Bricken, Tom Henighan, and Tristan Hume. Adly Templeton, in collaboration with Trenton Bricken and Tom Conerly, performed the engineering to scale up our experiments to large numbers of tokens. Robert Lasenby created tooling that allowed us to train extremely small language models. Yifan Wu imported the \"Pile\" dataset so we could train models on a standard external dataset. Tom Conerly broadly assisted with our infrastructure and maintaining high code quality. Adly Templeton implemented automatic plots comparing different experiments in a scan.\nSparse Autoencoders (Algorithms / ML)\n– Trenton initially implemented and advocated for sparse autoencoders as an approach to dictionary learning. Trenton Bricken and Adly Templeton then collaborated on the research needed to achieve our results, including scanning hyperparameters, iterating on algorithms, introducing the \"neuron resampling\" method, and characterizing the importance of dataset scale. (It's hard for the other authors to communicate just how much work Trenton Bricken and Adly Templeton put into iterating on algorithms and hyperparameters.) Josh Batson assisted by analyzing runs and designing metrics with Chris Olah, Adly Templeton and Trenton Bricken to measure success.\nAnalysis Infrastructure\n– The tooling to collect data for the visualizations was created by Trenton Bricken and Adly Templeton. Tom Conerly greatly accelerated this. Adly Templeton implemented the tooling to perform large scale feature ablation analysis, as well as creating the shuffled weights models as baselines.\nInterface\n– The interface for visualizing and exploring features was created by Brian Chen, with support from Shan Carter. It replaced a much earlier version by Trenton Bricken.\nAnalysis\nFeature Deep Dives\n– The dataset examples and ablations came from Brian Chen's interface work and Adly Templeton's ablation infrastructure. Chris Olah developed the activation spectrum plots, broken down by proxy. Josh Batson developed the logit scatter plot visualization and the sensitivity analysis. Tom Henighan created the pinned sampling visualizations. Nick Turner, Josh Batson and Tom Henighan explored how best to analyze neuron alignment. Many people contributed to a push to systematically analyze many features in detail, including Tom Henighan, Josh Batson, Trenton Bricken, Adly Templeton, Nick Turner, Brian Chen, and Chris Olah.\nManual Analysis of Features\n– Nick Turner and Adam Jermyn created our rubric for evaluating if a feature interval is interpretable. Adam Jermyn manually scored a random set of features.\nAutomated Interpretability\n– Our automated interpretability pipeline was created by Trenton Bricken, Cem Anil, Carson Denison, and Amanda Askell. Trenton Bricken ran the experiments using it to evaluate our features. This built on earlier explorations of automated interpretability by Shauna Kravec, and was only possible due to infrastructure contributions by Tim Maxwell, Nicholas Schiefer, and Nicholas Joseph.\nFeature Splitting\n– Josh Batson, Adam Jermyn, and Chris Olah analyzed feature splitting. Shan Carter produced the UMAPs of features.\nUniversality\n– Josh Batson analyzed universality, with engineering support from Trenton Bricken and Tom Henighan.\n\"Finite State Automata\"\n– Chris Olah analyzed \"Finite State Automata\".\nPaper Production\nWriting\n– The manuscript was primarily drafted by Chris Olah, Josh Batson, and Adam Jermyn, with extensive feedback and editing from all other authors. The detailed appendix on dictionary learning was drafted by Adly Templeton and Trenton Bricken. Josh Batson and Chris Olah managed the revision process in response to internal and external feedback.\nIllustration\n– Diagrams were created by Chris Olah, Josh Batson, Adam Jermyn, and Shan Carter, based on data generated by themselves and others.\nDevelopment of Intuition & Negative Results\nEarly Dictionary Learning Experiments\n– Many of the authors (including Trenton Bricken, Adly Templeton, Tristan Hume, Tom Henighan, Adam Jermyn, Josh Batson, and Chris Olah) did experiments applying both more traditional dictionary learning methods and sparse autoencoders to a variety of toy problems and small MNIST models. Much of this work focused on finding metrics which we hoped would tell us whether dictionary learning had succeeded in a simple context. These experiments, along with theoretical work, built valuable intuition and helped us discover several algorithmic improvements, and clarified challenges with traditional dictionary learning.\nSparse Architectures\n– Brian Chen ran the experiments and generated the counter-examples which persuaded us that training models for sparse activations was less promising. Crucially, he produced models where neurons typically activated in isolation, and showed they were still polysemantic. He then produced the counter-example described in the text. Robert Lasenby implemented infrastructure that made it easier to experiment with sparse activation architectures.\nOther\nSupport\n– Alex Tamkin, Karina Nguyen, Brayden McLean, and Josiah E Burke made a variety of contributions through infrastructure, operational support, labeling features, and commenting on the draft.\nLeadership\n– Tom Henighan led the dictionary learning project. Tristan Hume led an early version of the project, before handing it over to Tom Henighan. Shan Carter managed the overall team. Chris Olah provided general research guidance.\nAcknowledgments\nWe are deeply grateful to Martin Wattenberg, Neel Nanda, Hoagy Cunningham, David Lindner, Logan Smith, Steven Bills, William Saunders, Jonathan Marcus, Daniel Mossing, Nick Cammarata, Robert Huben, Aidan Ewart, Nicholas Sofroniew, Anna Golubeva, Bruno Olshausen for their detailed comments and feedback, which greatly improved our work.\nWe are also grateful to our colleagues at Anthropic who generously provided feedback, and many of whom also helped us explore the features discovered by dictionary learning as part of a \"Feature Party\". We particularly thank Oliver Rausch, Pujaa Rajan, Anna Chen, Alexander Silverstein, Marat Freytsis, Maryam Mortazavi, Mike Lambert, Justin Spahr-Summers, Emmanuel Ameisen, Andre Callahan, Shannon Yang, Zachary Witten, Zac Hatfield-Dodds, Karina Nguyen, Avital Balwit, Amanda Askell, Brayden McLean, and Nicholas Schiefer.\nOur work is only possible because of the extensive support of all our colleagues at Anthropic, from infrastructure, to engineering, to operations and more. It's impossible for us to list all the people whose work indirectly supported this paper, because there are so many, but we're deeply grateful for their support.\nCitation Information\nPlease cite as:\nBricken, et al., \"Towards Monosemanticity: Decomposing Language Models With Dictionary Learning\", Transformer Circuits Thread, 2023.\nBibTeX Citation:\n@article{bricken2023monosemanticity,\n       title={Towards Monosemanticity: Decomposing Language Models With Dictionary Learning},\n       author={Bricken, Trenton and Templeton, Adly and Batson, Joshua and Chen, Brian and Jermyn, Adam and Conerly, Tom and Turner, Nick and Anil, Cem and Denison, Carson and Askell, Amanda and Lasenby, Robert and Wu, Yifan and Kravec, Shauna and Schiefer, Nicholas and Maxwell, Tim and Joseph, Nicholas and Hatfield-Dodds, Zac and Tamkin, Alex and Nguyen, Karina and McLean, Brayden and Burke, Josiah E and Hume, Tristan and Carter, Shan and Henighan, Tom and Olah, Christopher},\n       year={2023},\n       journal={Transformer Circuits Thread},\n       note={https://transformer-circuits.pub/2023/monosemantic-features/index.html}\n    }\nFeature Proxies\nIn order to analyze features, we construct \"proxies\" for each feature which can be easily computed. Our proxies are the log-likelihood ratio of a string under the feature hypothesis and under the full empirical distribution. For example,\n\\log(P(s|\\text{base64}) / P(s))\nor\n\\log(P(s|\\text{Arabic Script}) / P(s))\n. We use log-likelihood proxies on the intuition that features, since they linearly interact with the logits, will be incentivized to track log-likelihoods.\nSince the activation of a feature at a token may be in part due to the preceding tokens (for example the Arabic feature could fire more strongly towards the end of a long Arabic string), we maximize the log-likelihood ratio over different prefixes leading up to and including the final token.\nBut how can we estimate the different terms? Let's consider each part separately.\nEstimating\nP(s)\nTo compute\nP(s)\n, we use the unigram distribution over tokens and compute\nP(s) = \\prod_{t\\in s} p(t)\n.\nEstimating\nP(s|\\text{base64})\n(and other characters features)\nFor base64, we assume that we're modeling a random base64 string where the probability of a character\nc\nis\nP(c) = 1/64\nif it's in a base64 character, and\n10^{-10}\notherwise. We then compute\nP(s) = \\prod_{c\\in s} P(c)\n.\nWe model DNA as a random string of\n[ATCG]\n, each with probability 1/4.\n(Although we don't use them in this draft, we also found this approach helpful for hexadecimal, binary, and similar features.)\nEstimating\nP(s|\\text{Arabic})\n(and other scripts features)\nWe also construct proxies for languages written in scripts that use distinctive unicode blocks. We'll use Arabic as an example of this.\nFor\nP(s|\\text{Arabic})\n, we use Bayes rule (i.e.\nP(s|\\text{Arabic}) = P(\\text{Arabic}|s) \\cdot P(s) / P(\\text{Arabic})\n). For\nP(s)\n, we use the method discussed above. For the other terms, we exploit the fact that it is easy to detect whether a character is in a Unicode block associated with Arabic script (including e.g.\nU+0600–U+06FF\n,\nU+0750–U+077F\n). Note this will count common characters shared between languages, like punctuation, as non-Arabic characters. If a string s consists entirely of Arabic characters,\nP(Arabic|s)\nis 1; if it contains non-Arabic characters,\nP(\\text{Arabic}|s)\nis assigned a tiny probability of 1e−10. (Keep in mind that we will be maximizing over multiple prefix strings.) If\ns\nconsists of a single token which isn't a complete unicode character, we sample random occurrences of it and observe the fraction of the time it is used in a given script. We also estimate\nP(\\text{Arabic})\nby applying this heuristic to a random sample of the dataset.\nFeature Ablations\nWe perform feature ablations by running the model on an entire context up through the MLP layer, running the autoencoder to compute feature activations, subtracting the feature direction times its activation from the MLP activation on each token in the context (replacing\n\\mathbf{x}^j\nwith\n\\mathbf{x}^j - f_i(\\mathbf{x}^j)\\mathbf{d}_j\n) and then completing the forward pass. We record the resulting change in the predicted log-likelihood of each token in the context in the color of an underline of that token. Thus if a feature were active on token [B] in the sequence [A][B][C], and ablating that feature reduced the odds placed on the prediction of C, then there would be an orange background on [B] (the activation) and a blue underline on [C] (the ablation effect), indicating that ablating that feature increased the model’s loss on the prediction of [C] and hence that feature is responsible for improving the model’s ability to predict [C] in that context.\nFeature Interpretability Rubric\nOur rubric for scoring how interpretable features are has the following instructions:\nForm an interpretation of the feature based on the positive logits and feature activations across all intervals.\nOn a scale of 0–3, rate your confidence in this interpretation.\nOn a scale of 0–5, rate how consistent the high-activation (bolded) tokens are with your interpretation from (1).\nOn a scale of 0–3, rate how consistent the positive logit effects are with your interpretation from (1).\nIf some of the positive logit effects were inconsistent with your interpretation, was there a separation in effect size between the consistent and inconsistent ones? If so, score as 1, otherwise or if not applicable, score as 0.\nOn a scale of 0–3, how specific is your interpretation of this feature?\nThe total score for a feature’s interpretability is the sum of ratings across these steps. Note that the maximum score is 14, because a perfect score on item #4 implies that item #5 is not applicable.\nFeature Density Histograms\nWhile iterating on different techniques, one important proxy for autoencoder performance is\nfeature density\n. Each feature in our autoencoder only activates on a very small percentage of the total tokens in the training set. We define the\nfeature density\nof each feature as the fraction of tokens on which the feature has a nonzero value.\nWe hypothesize that language models contain a large number of features across a distribution of feature densities, and that lower-density features are harder for our autoencoder to discover because they appear less often in the training dataset. Using large training datasets was an attempt to recover such low-density features.\nAs one way to measure the performance of an autoencoder, we take all the feature densities and plot a histogram of their distribution on a log scale. Roughly, we look for two histogram metrics: the number of features we’ve recovered, and the minimum feature density among features that we’ve recovered. Anecdotally, we find that these metrics are a decent proxy for subjective autoencoder performance and useful for quick iteration cycles.\nIdeally, the autoencoder neurons would be either meaningful features or completely dead. Indeed, training does completely leave some autoencoder neurons unused. Many neurons, however, seem to fall into a third category, which we tentatively name the\nultralow density cluster\n.\nMany of the feature density histograms are bimodal. For example, consider the feature density histogram for A/16:\nThis histogram has two main modes: a left mode around\n10^{-7}\n(corresponding to the ultralow density cluster) and a right mode around\n10^{-5}\n.\nThe right mode in this example may actually be split into two modes, one around\n10^{-5}\nand another around\n10^{-3}\n. We are still investigating the nature of this split.\nWe call the right mode the “high density cluster”, although the absolute magnitude of the density is very low. Anecdotally, almost all of the features in the high density cluster are interpretable, but almost none of the features in the ultralow density cluster are. As an aside, we suspect that this bimodality in feature density might partially explain the bimodality in cross-dictionary MCS observed by Huben\n.\nIn many preliminary small-scale experiments, the two clusters are often perfectly separable. However, for runs with a large number of autoencoder neurons and a large L1 coefficient these two clusters often overlap.\nEven in these cases, the two clusters can be separated by combining other statistics. For each feature, in addition to density, we plot the bias and the dot product of the vectors in the decoder and the encoder corresponding to each feature. Along these three dimensions, the high density cluster is clearly separable.\nThe ultralow density cluster appears to be an artifact of the autoencoder training process and not a real property of the underlying transformer. We are continuing to investigate the source of this phenomenon and other ways to either mitigate these features or to robustly automatically detect and filter them.\nBelow, we present feature density histograms for all Seed A runs, grouped by L1 coefficient. As we increase the number of autoencoder neurons (or “N learned sparse”), we continue to discover not only more features but also rarer features, suggesting that we have not exhausted the features in our one-layer transformer.\nTransformer Training and Preprocessing\nThe one-layer transformers we study are trained on the Pile\n. We chose to train these models on the Pile over Anthropic's internal dataset in order to make our experiments more reproducible. While we think many features are universal, others are very likely idiosyncratic to the dataset.\nWe train the transformers on 100 billion tokens using the Adam optimizer\n. We hypothesize that a very high number of training tokens may allow our model to learn cleaner representations in superposition. These transformers have a residual stream dimension of 128, and an inner MLP dimension of 512. The MLP activation is a ReLU.\nAdvice for Training Sparse Autoencoders: Autoencoder Dataset and Basic Training\nTo create the dataset for autoencoder training, we evaluate the transformers on 40 million contexts from the Pile and collect the MLP activation vectors after the ReLU for each token within each context. We then sample activation vectors from 250 tokens in each context and shuffle these together so that samples within a batch come from diverse contexts.\nFor training, we sample MLP activation vectors without replacement using a batch size of 8192. We’ve found that sampling without replacement (i.e., not repeating data) is important for optimal results. We perform 1 million update steps, using just over 8 billion activation vectors out of the total dataset size of 10 billion. We use an Adam optimizer\nto minimize the sum of mean squared error loss and an L1 regularization penalty on the hidden layer activations of the autoencoder.\nAdvice for Training Sparse Autoencoders: Autoencoder Architecture\nOur dictionary learning model is a one hidden layer MLP. It is trained as an autoencoder, using the input weights as an encoder and output weights as the decoder. The hidden layer is much wider than the inputs and applies a ReLU non-linearity. We use the default Pytorch Kaiming Uniform initialization\n.\nFormally, let\nn\nbe the input and output dimension and\nm\nbe the autoencoder hidden layer dimension. Given encoder weights\nW_e \\in \\mathbb{R}^{m \\times n}\n, decoder weights\nW_d \\in \\mathbb{R}^{n \\times m}\nwith columns of unit norm, and biases\n\\mathbf{b}_e \\in \\mathbb{R}^{m}, \\mathbf{b}_d \\in \\mathbb{R}^{n}\n, the operations and loss function over a dataset\nX\nare:\n\\begin{aligned} \\bar{\\mathbf{x}} &= \\mathbf{x} - \\mathbf{b}_{d} \\\\ \\mathbf{f} &= \\text{ReLU}( W_e \\bar{\\mathbf{x}}+\\mathbf{b}_e ) \\\\ \\hat{\\mathbf{x}} &= W_d \\mathbf{f}+\\mathbf{b}_d \\\\ \\mathcal{L} &= \\frac{1}{|X|} \\sum_{\\mathbf{x}\\in X}  ||\\mathbf{x}-\\hat{\\mathbf{x}}||_2^2 + \\lambda||\\mathbf{f}||_1  \\end{aligned}\nNote that our hidden layer is overcomplete\nm\\geq n\n, and the MSE is a mean over each vector element while the L1 penalty is a sum (with\n\\lambda\nbeing the L1 coefficient). The hidden layer activations\n\\mathbf{f}\nare our learned features. As shown, we subtract the decoder bias from the inputs, and call this a pre-encoder bias. Recall that the decoder is also referred to as the “dictionary” while the encoder can be thought of as a linear, amortized approximation to more powerful sparse coding algorithms.\nOur modifications to a standard autoencoder are backed by theory and empirical ablations. However, because these modifications appeared at different periods during our research and can interact with each other, we do not provide rigorous ablation experiments and instead provide intuition and anecdotal evidence for why each is justified. There are likely many important modifications that could further improve performance.\nHere is a summary of the modifications sorted both between and within sections in rough order of importance\nPre encoder bias – this boosted performance in toy models.\nDecoder weights not tied – the learned encoder weights are often far from the corresponding decoder weights and this is important for increasing model capacity.\nNeuron resampling – helps finding more features and achieve a lower total loss.\nMany training steps – beyond when the training loss looks to have plateaued.\nLearning rate sweep – lower learning rate results in more real features.\nInteraction Between Adam and Decoder Normalization – a principled way to account for how the dictionary vector L2 normalization interacts with Adam, resulting in lower total loss.\nPre-Encoder Bias\nTesting the autoencoder on toy problems with a bias, we found that it would fail to produce the “bounce plots” we\ndescribed previously\nunless we added a learned bias term to both the decoder output and the encoder input. We constrain the pre-encoder bias to equal the negative of the post-decoder bias and initialize it to the geometric median of the dataset. These investigations were motivated by the observations by Hobbhahn\n.\nDecoder Weights Not Tied\nIt is common to tie the encoder and decoder weights of single hidden layer autoencoders\n. However, we find that in our trained models the learned encoder weights are\nnot\nthe transpose of the decoder weights and are cleverly offset to increase representational capacity. Specifically, we find that similar features which have closely related dictionary vectors have encoder weights that are offset so that they prevent crosstalk between the noisy feature inputs and confusion between the distinct features.\nFor example, in\n\"Bug\" 2: Multiple Features for a Single Context\nwe described 3 features handling different types of base64 strings. These all have similar dictionary vectors, but it turns out that the encoder weight vectors, while still similar, are more tilted away to help distinguish them.\nOne way of viewing this is that our autoencoder is learning an amortized, linear approximation to a multi-step, non linear sparse coding algorithm. Allowing the weights of the encoder to be independent from the decoder enables more representational capacity as we empirically observe.\nNeuron Resampling\nOver the course of training, a subset of autoencoder neurons will have zero activity across a large number of datapoints. We find that “resampling” these dead neurons during training improves the number of likely-interpretable features (i.e., those in the high density cluster, see\nFeature Density Histograms\n) and reduces total loss. This resampling may be compatible with the Lottery Ticket Hypothesis\nand increase the number of chances the network has to find promising feature directions.\nAn interesting nuance around dead neurons involves the ultralow density cluster. We find that if we increase the number of training steps then networks will kill off more of these ultralow density neurons. This reinforces the use of the high density cluster as a useful metric because there can exist neurons that are de facto dead but will not appear to be when looking at the number of dead neurons alone.\nThe number of dead neurons that appear over training appears to depend upon a number of factors including but not limited to: learning rate (too high); batch size (too low); dataset redundancy (too many tokens per context or repeated epochs over the same dataset); number of training steps (too many); optimizer used. Our results here agree with those of Bricken\net al.\n.\nBetter resampling strategies is an active research area of ours. The approach used in this work is as follows:\nAt training steps 25,000, 50,000, 75,000 and 100,000, identify which neurons have not fired in any of the previous 12,500 training steps.\nCompute the loss for the current model on a random subset of 819,200 inputs.\nAssign each input vector a probability of being picked that is proportional to the square of the autoencoder’s loss on that input.\nFor each dead neuron sample an input according to these probabilities. Renormalize the input vector to have unit L2 norm and set this to be the dictionary vector for the dead autoencoder neuron.\nFor the corresponding encoder vector, renormalize the input vector to equal the average norm of the encoder weights for alive neurons × 0.2. Set the corresponding encoder bias element to zero.\nReset the Adam optimizer parameters for every modified weight and bias term.\nThis resampling procedure is designed to seed new features to fit inputs where the current autoencoder performs worst. Resetting the encoder norm and bias are crucial to ensuring this resampled neuron will only fire weakly for inputs similar to the one used for its reinitialization. We do this to minimize interference with the rest of the network.\nThis resampling approach outperforms baselines including no resampling and reinitializing the relevant encoder and decoder weights using a default Kaiming Uniform initialization. However, there are likely better resampling methods to be developed – this approach still causes sudden loss spikes, and resampling too frequently causes training to diverge.\nLearning Rate Sweep\nWe performed several training runs using multiple learning rates while also varying the number of training steps and found that lower learning rates, when given sufficient training steps, result in lower total loss and more “real” features discovered (as indicated by the\nFeature Density Histograms\n). We also tried annealing our learning rate over the course of training but found that this did not further increase performance.\nInteraction Between Adam and Decoder Normalization\nRecall that we constrain our dictionary vectors to have unit norm. Our first naive implementation simply reset all vectors to unit norm after each gradient step. This means any gradient updates modifying the length of our vector are removed, creating a discrepancy between the gradient used by the Adam optimizer and the true gradient. We find that instead removing any gradient information parallel to our dictionary vectors before applying the gradient step results in a small but real reduction in total loss.\nAdvice for Training Sparse Autoencoders: Hyperparameter Selection\nThe ultimate goal of dictionary learning is to produce features that are interpretable and also accurately represent the underlying model. Quantitatively measuring interpretability is currently difficult and slow, so we frequently look at proxy metrics when testing changes or optimizing hyperparameters. Our methods for optimizing hyperparameters are far from perfect, but we want to share insights that may be useful for other researchers doing similar work.\nWe most frequently look at the following proxy metrics:\nTraining loss:\nThis is the simplest metric, and useful for testing changes such as optimizers or learning rates. Unfortunately, it isn’t meaningful to compare training loss across hyperparameters that change the loss function, such as L1 coefficients.\nFeature Density Histograms\n:\nSpecific metrics from these histograms include:\nThe number of alive features outside of the ultralow density cluster\nThe minimum feature density at which we see a significant number of non-ultralow-density-cluster features.\nThe number of features with density above 1%. A significant number of features above this level seems to correspond to an L1 coefficient that is too low.\n\\mathbf{L^0}\nnorm:\nThe average number of nonzero entries in a sparse representation. We don’t yet know of a principled way to determine a target level of sparsity, but we generally target a\nL^0\nnorm that is less than 10 or 20. We especially distrust solutions where the\nL^0\nnorm is a significant fraction of the transformer’s activation dimensionality.\nReconstructed Transformer NLL:\nWe would like the features we discover to explain almost all of the behavior of the underlying transformer. One way to measure this is to take a transformer, run the MLP activations through our autoencoder, replace the MLP activations with the autoencoder predictions, measure the loss on the training dataset, and calculate the difference in loss.\nWe often normalize this by dividing by the difference in loss between the baseline transformer’s performance and its performance after ablating the MLP layer. This gives us a fraction of the MLP’s loss contribution that is explained by our transformer. However, the performance with an ablated MLP may be an especially bad baseline, so this percentage is considered an overestimate. Ad hoc experiments (data not shown) show that similarly high percentages can be reached by training a transformer from scratch with a small MLP.\nThe\nL^0\nnorm and Reconstructed Transformer NLL metrics, taken together, display a human-interpretable tradeoff between sparsity and explained behavior:\nEach line on this plot shows data for a single L1 coefficient and varying autoencoder sizes. The lowest L1 coefficient is excluded to create a reasonable scale.\nActivation Visualization\nVisualizations were produced using a 100 million token subset of our training dataset where only 10 tokens are taken from every context. We display a window of four tokens on each side of the selected token for visualization.\nFeature UMAPs\nWe construct two UMAP embeddings of learned features from our transformer. We represent each feature as a vector of dimension 512 (the number of neurons), its column in the decoder matrix\nW_d\n. The first UMAP embeds the features from A/0 and A/1 jointly into 2 dimensions, and uses hyperparameters\nn_neighbors=15\n,\nmetric=\"cosine\"\n,\nmin_dist=0.05\n. The second UMAP uses the same hyperparameters, but with\nmin_dist=0.01\n, and also includes the features from A/2. We additionally organize the points in the first UMAP into clusters by applying HDBSCAN with\nmin_cluster_size=3\nto a 10-dimensional UMAP embedding of the same features, fit with hyperparameters\nn_neighbors=15\n,\nmetric=\"cosine\"\n,\nmin_dist=0.1\n. The list of features is sorted by a one-dimensional UMAP fit to the unclustered features and the cluster mediods.\nAutomated Interpretability\nHere we provide more details on the implementation of our automated interpretability before explaining importance scoring and providing additional results.\nExperimental Setup\nIn order to get feature explanations Claude 2 is provided with a total of 49 examples: ten examples from the top activations interval; two from the other 12 intervals; five completely random examples; and ten examples where the top activating tokens appear in different contexts.\nWe are explicit in keeping all of these examples separate from the ones used when we test on activation predictions.\nFor example, in the main text we discuss a feature that fires for token \"the\" in the context of machine learning, the ten examples here would contain the word \"the\" but across different contexts. We also provide the top and bottom logits and labels for what interval each example came from. In addition, we give eight human demonstrations on held out features that include chain of thought reasoning. Following Bills\net al.\nall of our activations are quantized to be between 0–9 inclusive. Finally, we ask the model to be succinct in its answer and not provide specific examples of tokens it activates for.\nUsing the explanation generated, in a new interaction Claude is asked to predict activations for sixty examples: six from the top activations; two from the other 12 intervals; ten completely random; and twenty top activating tokens out of context. The same eight held out demonstration features show how each token should elicit a predicted activation. All of the examples are shuffled, the logits and interval each came from is removed, and blanks are left where the model should fill in its activation predictions. For the sake of computational efficiency, Claude scores all sixty examples in a single shot, repeating each token followed by its predicted activation. In an ideal setting, each example would be given independently as its own prompt.\nNote that there are instances where a feature has less than 49 explanation examples and 60 prediction examples because we remove any repeated examples that appear in the feature visualization intervals. However, this occurs for a negligible number of features and we drop any of those with fewer than 20 examples in total.\nFor the sake of comparison to Bills\net al.\nthe figure below includes Pearson correlation in addition to Spearman that is otherwise used throughout the text. The distributions and overall conclusions are almost identical to Spearman.\nWe also systematically catalog the feature output weight logits with Claude. We use the same explanation Claude generated to predict activations and provide 100 logits, 50 taken randomly and 50 which are the 10th–60th largest positive logits (the top 10 were used to generate the original explanations and thus held out). Claude was asked to output a binary label for if this logit would plausibly come after the feature fired, given the explanation. Note that the chance performance is 50%. The A/1 features are quite a bit better than the neurons at a median of 0.74 versus 0.53 for the neurons and 0.5 for the randomized transformer.\nRandomized Transformer\nAs a further ablation, we compare our human and automated interpretability approaches to the randomized transformer. It is unsurprising that the logit weight predictions are centered around random chance for the randomized transformer by the very nature of its random weights. Manual interpretability also favors the A/1 features. However, for automated interpretability the randomized features have a higher median score.\nIt turns out that the randomized transformer features are distinctly bimodal, consisting of single token and polysemantic clusters which are separated out in the inset at the bottom of the following figure.\nFeatures are labeled as being \"single token\" if for all 20 examples in the top activation interval, the token with the largest activation is the same. The single token features of the randomized transformer are not only more numerous but also more truly single token, for example, firing for the same token even in the weakest activation intervals. This makes them more trivial to score as evidenced by the high scoring yellow peak (bottom row left side of the figure).\nMeanwhile, for the not \"single token\" cluster, the A/1 features score higher while the randomized features in fact have a similar distribution to that of the polysemantic neurons.\nNote that this bimodal clustering of the randomized transformer features is also supported by the manual human analysis where there is both a cluster at 0, the same score given to the polysemantic neurons in the main text, and a non zero cluster. The non-zero cluster scores worse than the A/1 Features here (unlike for the automated approach) because part of the rubric accounts for the logit interpretability, which scores zero in the shuffled case.\nThe fact that the randomized transformer learns more highly single token features fits with the hypothesis that dictionary learning will learn the features of the dataset rather than the representations of the transformer.\nImportance Scoring\nImportance scoring is the principled way to correct for the fact that we did not randomly sample all of the examples our model is asked to score. Formally, we assign to all tokens from random examples a weight of 1 and examples from each feature interval a weight of\nfeature_density*interval_probability\n.\nfeature_density\nis the fraction of times this feature fires over a large portion of the dataset and\ninterval_probability\nconverts the distribution of non-zero activations into a categorical distribution corresponding to each interval and uses the corresponding probability. We then use these weights in our Spearman correlation as before.\nThe top activating tokens out of context examples inherit the weighting of the top activating examples quantile.\nThe issue with importance scoring is that because our features have densities typically in the range of 1e−3 to 1e−6  (see next Appendix section), almost all of the weight is assigned to the random examples for which features almost never fire. As a result, it is possible for features to score well by almost always predicting zero and in turn favors explanations that are overly restrictive that say what the feature does\nnot\nfire for, instead of what it does fire for. Conversely, explanations are highly penalized for even a small false positive rate, since the true rate of occurrences is so low.\nThe figure below shows importance scores for features and neurons. Note that many of the features now have scores very close to zero while the neurons are relatively less affected. We hypothesize this is because the neurons, in being more polysemantic, have more non-zero activations that test their explanations on both things it does and doesn't fire for. Empirically, taking an average over features, 88% of their true activations from the random examples are zero versus 54% for the neurons (the medians are ~91% and 58%, respectively).\nExplanation Lengths Caveat\nClaude can at times fail to wrap its explanation in the\n<answer></answer>\nexplanation tags resulting in the full chain of thought being used instead. This can be much longer and leak more specific words that the feature fires for.\nHere we plot the number of characters in each explanation where the second mode to the right highlights these instances. However, not only does this happen infrequently, it also occurs across the features, neurons, and randomized transformer results. In fact, it happens the most often for neurons, occurring 19% of the time versus 15% for the randomized transformer and 13% for the features. Most importantly, it does not result in significantly different scores when we split out the performance of those with long explanations from short ones.\nTo conclude, while automated interpretability is a difficult task for models that we have only begun working on, it has already been very useful for quickly understanding dictionary learning features in a scalable fashion. We encourage readers to explore the visualizations with automated interpretability scores:\nA/1\n,\nA/0\n,\nrandomized model features\n,\nneurons\n.",
  "metadata": {
    "title": "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning",
    "description": null,
    "keywords": null,
    "author": null
  },
  "paywall": {
    "listed_paid": false,
    "detected": false,
    "marker": "",
    "public_only": false,
    "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙"
  },
  "crawl": {
    "success": true,
    "status_code": 200,
    "error": ""
  },
  "code_blocks": [
    {
      "file": "attachments/37d897fcb626/code/body-000.txt",
      "chars": 132,
      "language": "text"
    },
    {
      "file": "attachments/37d897fcb626/code/body-001.txt",
      "chars": 775,
      "language": "text"
    }
  ],
  "images": [],
  "audio": [],
  "videos": [],
  "links": [
    {
      "href": "https://anthropic.com/",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/",
      "label": "Transformer Circuits Thread",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html",
      "label": "Browse A/1 Features →",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/index.html",
      "label": "Browse All Features →",
      "type": "a"
    },
    {
      "href": "https://www.anthropic.com/",
      "label": "Anthropic",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/colah@anthropic.com",
      "label": "colah@anthropic.com",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?ordering=index",
      "label": "a single neuron",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2022/toy_model/index.html",
      "label": "Toy Models of Superposition",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/superposition-composition/index.html",
      "label": "Distributed Representations: Composition and Superposition",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/",
      "label": "90 learned dictionaries",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/july-update/index.html",
      "label": "safety-relevant features",
      "type": "a"
    },
    {
      "href": "https://en.wikipedia.org/wiki/Sparse_dictionary_learning",
      "label": "sparse dictionary learning",
      "type": "a"
    },
    {
      "href": "https://en.wikipedia.org/wiki/Compressed_sensing",
      "label": "compressed sensing",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/may-update/index.html",
      "label": "\"too strong\"",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html?search_target_mode=index_name&search_mode=case_insensitive",
      "label": "fantasy games",
      "type": "a"
    },
    {
      "href": "https://www.compart.com/en/unicode/U+062B",
      "label": "U+062B",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a1-ar.html",
      "label": "alternative visualization",
      "type": "a"
    },
    {
      "href": "https://www.compart.com/en/unicode/block/U+0600",
      "label": "basic Arabic Unicode block",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?search_mode=regex_case_sensitive&ordering=count&search_text=%5B%5Cu0600-%5Cu06FF%5D+",
      "label": "search",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html",
      "label": "A/neurons/489",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html",
      "label": "B/1/1334",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a1-dna.html",
      "label": "but by the end of a long DNA sequence",
      "type": "a"
    },
    {
      "href": "https://en.wikipedia.org/wiki/Base64",
      "label": "base64",
      "type": "a"
    },
    {
      "href": "https://www.compart.com/en/unicode/block/U+0590",
      "label": "basic Hebrew Unicode block",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?search_mode=regex_case_sensitive&ordering=count&search_text=%5B%5Cu0590-%5Cu05FF%5D",
      "label": "searched",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html?ordering=max_dense",
      "label": "most active features",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a5.html",
      "label": "A/",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/random1.html",
      "label": "here",
      "type": "a"
    },
    {
      "href": "https://en.wikipedia.org/wiki/Phenomenology_(physics)",
      "label": "phenomenology in physics",
      "type": "a"
    },
    {
      "href": "https://deep-phenomena.org/",
      "label": "2019 ICML workshop",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html",
      "label": "A/0/341",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a4.html",
      "label": "A/4",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a2.html",
      "label": "A/2/12310",
      "type": "a"
    },
    {
      "href": "https://distill.pub/2020/circuits/zoom-in/",
      "label": "universal",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2022/solu/index.html",
      "label": "Section 6.3.1",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/monosemantic-features/vis/a3.html",
      "label": "A/3/16085",
      "type": "a"
    },
    {
      "href": "https://www.compart.com/en/unicode/block/U+0B80",
      "label": "block U+0B80–U+0BFF",
      "type": "a"
    },
    {
      "href": "https://www.compart.com/en/unicode/U+0BA3",
      "label": "U+0BA3",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/abs/1601.03764",
      "label": "Arora",
      "type": "a"
    },
    {
      "href": "https://transformer-circuits.pub/2023/interpretability-dreams/index.html",
      "label": "here",
      "type": "a"
    },
    {
      "href": "https://distill.pub/2020/circuits/",
      "label": "Circuits Thread",
      "type": "a"
    },
    {
      "href": "https://distill.pub/2019/advex-bugs-discussion/",
      "label": "Distill's Discussion Article experiment",
      "type": "a"
    },
    {
      "href": "https://www.alignmentforum.org/posts/fKuugaxt2XLTkASkk/open-source-replication-and-commentary-on-anthropic-s",
      "label": "a blog post",
      "type": "a"
    },
    {
      "href": "https://colab.research.google.com/drive/1u8larhpxy8w4mMsJiSBddNOzFGj7_RTn?usp=sharing",
      "label": "an accompanying tutorial",
      "type": "a"
    },
    {
      "href": "https://github.com/neelnanda-io/1L-Sparse-Autoencoder",
      "label": "training codebase",
      "type": "a"
    },
    {
      "href": "https://www.compart.com/en/unicode/block/U+0750",
      "label": "U+0750–U+077F",
      "type": "a"
    }
  ],
  "repair_version": 2,
  "repaired_at": "2026-10-20T17:24:04Z",
  "source_url": "https://transformer-circuits.pub/2023/monosemantic-features/",
  "delivery_status": "ok",
  "discovery_status": "ok",
  "missing_reason": "",
  "repair_error": "",
  "actual_video_files": [
    {
      "local_path": "attachments/37d897fcb626/videos/dQw4w9WgXcQ.f251.webm",
      "bytes": 3433755,
      "valid": false,
      "height": 0,
      "width": null,
      "duration": "213.061000",
      "status": "invalid_or_wrong_resolution"
    },
    {
      "local_path": "attachments/37d897fcb626/videos/dQw4w9WgXcQ.f399.mp4",
      "bytes": 30415996,
      "valid": true,
      "height": 1080,
      "width": 1920,
      "duration": "213.040000",
      "status": "ok"
    }
  ],
  "expected_saved_body_images": 2,
  "image_discovery": "saved_markdown",
  "source_body_markdown": "[ ](https://anthropic.com) [Transformer Circuits Thread](https://transformer-circuits.pub/)\n# Towards Monosemanticity: Decomposing Language Models With Dictionary Learning\n# Towards Monosemanticity: Decomposing Language Models With Dictionary Learning\nUsing a sparse autoencoder, we extract a large number of interpretable features from a one-layer transformer.  \n[Browse A/1 Features →](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html)  \n[Browse All Features →](https://transformer-circuits.pub/2023/monosemantic-features/vis/index.html)\n### Authors\nTrenton Bricken*, Adly Templeton*, Joshua Batson*, Brian Chen*, Adam Jermyn*, Tom Conerly, Nicholas L Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, Chris Olah\n### Affiliations\n[Anthropic](https://www.anthropic.com/)\n### Published\nOct 4, 2023\n* Core Contributor; Correspondence to [colah@anthropic.com](https://transformer-circuits.pub/2023/monosemantic-features/colah@anthropic.com); [Author contributions statement below](https://transformer-circuits.pub/2023/monosemantic-features/#author-contributions). \n### Authors\n### Affiliations\n### Published\n_Not published yet._\n### DOI\n_No DOI yet._\n### Contents\n[Problem Setup](https://transformer-circuits.pub/2023/monosemantic-features/#problem-setup)\n  * [Features as a Decomposition](https://transformer-circuits.pub/2023/monosemantic-features/#problem-setup-decomposition)\n  * [What makes a good decomposition?](https://transformer-circuits.pub/2023/monosemantic-features/#problem-setup-good-decomposition)\n  * [Why not use architectural approaches?](https://transformer-circuits.pub/2023/monosemantic-features/#why-not-architectures)\n  * [Using Sparse Autoencoders To Find Good Decompositions](https://transformer-circuits.pub/2023/monosemantic-features/#setup-autoencoder-motivation)\n  * [Sparse Autoencoder Setup](https://transformer-circuits.pub/2023/monosemantic-features/#setup-autoencoder)\n  * [The (one-layer) model we're studying](https://transformer-circuits.pub/2023/monosemantic-features/#setup-transformer)\n  * [Notation for Features](https://transformer-circuits.pub/2023/monosemantic-features/#setup-notation)\n\n\n[Detailed Investigations of Individual Features](https://transformer-circuits.pub/2023/monosemantic-features/#feature-analysis)\n  * [Arabic Script Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic)\n  * [DNA Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-dna)\n  * [base64 Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-base64)\n  * [Hebrew Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-hebrew)\n\n\n[Global Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis)\n  * [How Interpretable is the Typical Feature?](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp)\n  * [How much of the model does our interpretation explain?](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-how-much)\n  * [Do features tell us about the model or the data?](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-about-model)\n\n\n[Phenomenology](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology)\n  * [Feature Motifs](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-motifs)\n  * [Feature Splitting](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-splitting)\n  * [Universality](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-universality)\n  * [\"Finite State Automata\"](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-fsa)\n\n\n[Related Work](https://transformer-circuits.pub/2023/monosemantic-features/#related-work)\n[Discussion](https://transformer-circuits.pub/2023/monosemantic-features/#discussion)\n[Comments & Replications](https://transformer-circuits.pub/2023/monosemantic-features/#comments)\n  * [Neel Nanda](https://transformer-circuits.pub/2023/monosemantic-features/#comment-nanda)\n\n\n[Author Contributions Statement](https://transformer-circuits.pub/2023/monosemantic-features/#author-contributions)\n[Acknowledgments](https://transformer-circuits.pub/2023/monosemantic-features/#acknowledgments)\n[ Citation Information](https://transformer-circuits.pub/2023/monosemantic-features/#citation)\n[Feature Proxies](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-proxies)\n[Feature Interpretability Rubric](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-rubric)\n[Feature Density Histograms](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-density)\n[Transformer Training and Preprocessing](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-transformer)\n[Advice for Training Sparse Autoencoders](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-dataset)\n[Automated Interpretability](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-automated)\nMechanistic interpretability seeks to understand neural networks by breaking them into components that are more easily understood than the whole. By understanding the function of each component, and how they interact, we hope to be able to reason about the behavior of the entire network. The first step in that program is to identify the correct components to analyze. \nUnfortunately, the most natural computational unit of the neural network – the neuron itself – turns out not to be a natural unit for human understanding. This is because many neurons are polysemantic: they respond to mixtures of seemingly unrelated inputs. In the vision model Inception v1, a single neuron responds to faces of cats and fronts of cars . In a small language model we discuss in this paper, [a single neuron](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?ordering=index#feature-83) responds to a mixture of academic citations, English dialogue, HTTP requests, and Korean text. Polysemanticity makes it difficult to reason about the behavior of the network in terms of the activity of individual neurons.\nOne potential cause of polysemanticity is superposition , a hypothesized phenomenon where a neural network represents more independent \"features\" of the data than it has neurons by assigning each feature its own linear combination of neurons. If we view each feature as a vector over the neurons, then the set of features form an overcomplete linear basis for the activations of the network neurons. In our previous paper on [Toy Models of Superposition](https://transformer-circuits.pub/2022/toy_model/index.html#strategic-ways-out) , we showed that superposition can arise naturally during the course of neural network training if the set of features useful to a model are sparse in the training data. As in compressed sensing, sparsity allows a model to disambiguate which combination of features produced any given activation vector.For more discussion of this point, see [Distributed Representations: Composition and Superposition](https://transformer-circuits.pub/2023/superposition-composition/index.html).\nIn Toy Models of Superposition, we described [three strategies](https://transformer-circuits.pub/2022/toy_model/index.html#strategic-ways-out) to finding a sparse and interpretable set of features if they are indeed hidden by superposition: (1) creating models without superposition, perhaps by encouraging activation sparsity; (2) using dictionary learning to find an overcomplete feature basis in a model exhibiting superposition; and (3) hybrid approaches relying on a combination of the two. Since the publication of that work, we've explored all three approaches. We eventually developed counterexamples which persuaded us that the sparse architectural approach (approach 1) was insufficient to prevent polysemanticity, and that standard dictionary learning methods (approach 2) had significant issues with overfitting.\nIn this paper, we use a weak dictionary learning algorithm called a sparse autoencoder to generate learned features from a trained model that offer a more monosemantic unit of analysis than the model's neurons themselves. Our approach here builds on a significant amount of prior work, especially in using dictionary learning and related methods on neural network activations (e.g. ), and a more general allied literature on disentanglement. We also note interim reports which independently investigated the sparse autoencoder approach in response to Toy Models, culminating in the recent manuscript of Cunningham et al. .\nThe goal of this paper is to provide a detailed demonstration of a sparse autoencoder compellingly succeeding at the goals of extracting interpretable features from superposition and enabling basic circuit analysis. Concretely, we take a one-layer transformer with a 512-neuron MLP layer, and decompose the MLP activations into relatively interpretable features by training sparse autoencoders on MLP activations from 8 billion data points, with expansion factors ranging from 1× (512 features) to 256× (131,072 features). We focus our detailed interpretability analyses on the 4,096 features learned in one run we call A/1.\nThis report has four major sections. In [Problem Setup](https://transformer-circuits.pub/2023/monosemantic-features/#problem-setup), we provide motivation for our approach and describe the transformers and sparse autoencoders we train. In [Detailed Investigations of Individual Features](https://transformer-circuits.pub/2023/monosemantic-features/#feature-analysis), we offer an existence proof – we make the case that several features we find are functionally specific causal units which don't correspond to neurons. In [Global Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis), we argue that the typical feature is interpretable and that they explain a non-trivial portion of the MLP layer. Finally, in [Phenomenology](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology) we describe several properties of our features, including [feature-splitting](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-splitting), [universality](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-universality), and how they can form [\"finite state automata\"-like systems](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-fsa) implementing interesting behaviors.\nWe also provide three comprehensive visualizations of features. First, for all features from [90 learned dictionaries](https://transformer-circuits.pub/2023/monosemantic-features/vis/) we present activating dataset examples and downstream logit effects. We recommend the reader begin with the [visualization of A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html). Second, we provide a [data-oriented view](https://transformer-circuits.pub/2023/monosemantic-features/vis/index.html#example-texts), showing all features active on each token of 25 texts. Finally, we coembed all 4,096 features from A/1 and all 512 features from A/0 into the plane using UMAP to allow for interactive exploration of the space of features:\nCluster\nFeature\nCluster #49\nA/0/307\nThis feature fires for references to citations in scientific papers. It attends to the formatting of citations such as brackets, @ symbols, names of authors, dates and labels.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-307)\nA/0/311\nThis feature fires for reference citations in academic papers, specifically when it sees the [@ symbol that is commonly used.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-311)\nA/1/776\nYears in some citation notation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-776)\nA/1/1538\nCitations in a [@author] or [@authoryear] format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1538)\nA/1/1875\nMarkdown Citation (Predict year)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1875)\nA/1/2252\n\" [@\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2252)\nA/1/2237\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2237)\nCluster #42\nA/0/126\nThis feature seems to fire on section headings, specifically the word \"sec\" within Markdown section tags like {#sec:foo}.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-126)\nA/1/357\n\"ref\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-357)\nA/1/1469\n\"s\"/\"sec\" after \"{#\", section reference in some markup\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1469)\nA/1/3841\n\"Sec\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3841)\nA/1/3898\nSection number in {#SecX}\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3898)\nA/1/4083\n\" {#\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4083)\nA/1/2129\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2129)\nA/1/553\n\"](#\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-553)\nCluster #43\nA/0/8\nThis feature attends to text formatting markups such as references, figure captions, and table captions. It looks specifically for the \"ref-type\" attribute in markup tags.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-8)\nA/0/398\nThis feature attends to references to figures and tables.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-398)\nA/0/454\nThis feature fires on reference/bibliographic citations in LaTeX documents. It attends to the braces, \"ref\" tag, and \"type\" attribute that are commonly used to define citations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-454)\nA/1/35\n\"){\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-35)\nA/1/366\n\"type\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-366)\nA/1/945\n\"ref\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-945)\nA/1/1895\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1895)\nA/1/2176\n\"fig\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2176)\nCluster #46\nA/0/425\nThis feature attends to highlighting bibliography and reference markers in the text such as [1], (Table 1), [@B23], etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-425)\nA/1/703\nThis feature attends to bibliographic citations with brackets and @CIT numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-703)\nA/1/864\n\" \\\\\\\\[[@\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-864)\nA/1/868\n\"[@\" in bibliography citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-868)\nA/1/2217\n\".](\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2217)\nA/1/2777\n\" ([@\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2777)\nA/1/3120\nThis feature attends to bibliographic citations or references, firing most strongly for formatted citations with brackets, citation keys like [b23] and full references citing authors, titles and dates.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3120)\nA/1/3184\nIntroducing an acaemic reference format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3184)\nA/1/3698\n\"](#\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3698)\nA/1/51\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-51)\nCluster #48\nA/0/363\nThe feature attends to in-text citations with URL and DOI links.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-363)\nA/0/366\nThis feature fires on citations and reference markers such as [ ], ( ), etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-366)\nA/0/396\nThis feature fires for in-text references, citation links, and reference brackets/notations in scholarly articles, research papers, and academic journals.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-396)\nA/1/90\n\"1\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-90)\nA/1/204\n\"CR\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-204)\nA/1/220\nThis feature attends to formats of references and citations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-220)\nA/1/729\nDigits in bibliographies?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-729)\nA/1/1015\nReference IDs in parentheses in [](){} citation format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1015)\nA/1/1264\nMostly letter fragments of IDs in parentheses in [](){} citation format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1264)\nA/1/1316\nA specific form of medical journal document reference, syntax '![](journalname12345-1234){...', specifically the journalid12345-1234 part\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1316)\nA/1/1440\n\"g\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1440)\nA/1/1445\n\"ijms\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1445)\nA/1/2200\nThe feature fires on citation indices in scientific papers, such as [B13-ijms-20-04353-f001].\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2200)\nA/1/2280\n\"bib\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2280)\nA/1/2385\n\"bib\" in some bibliography format ending in \"bib-XXXX\" where X are digits\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2385)\nA/1/2496\n\"B\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2496)\nA/1/2559\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2559)\nA/1/2696\nNumbers/digits in hyphenated IDs in square brackets, likely [@ ... ] citation format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2696)\nA/1/2881\n\"B\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2881)\nA/1/2988\nDigits after upper case letters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2988)\nA/1/2996\nThis feature attends to inline citations like [R9] or [@ref9] in scientific papers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2996)\nA/1/3143\nThis feature seems to fire on citations and references in brackets and parentheses with numbers, IDs and abbreviations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3143)\nA/1/3163\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3163)\nA/1/3164\nNumbers after \"pone.\", for citation IDs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3164)\nA/1/3547\nBeginnings of hash locations in Markdown URL targets?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3547)\nA/1/3949\n\"pone\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3949)\nA/1/4062\nPost-\"#\" numeric ID fragments in parentheses in [](){} citation format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4062)\nA/0/270\nThis feature fires on reference links to articles, papers, and datasets in academic writing. These references links are typically in a specific format like [text](link) or with abbreviated journal titles and citation numbers like [B14-sensors-19-00441].\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-270)\nCluster #36\nA/0/277\nThe feature fires when it sees short HTML/XML tags such as <ref-type=\"fig\">\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-277)\nA/0/427\nThis feature appears to fire on hyphens between numbers, especially years.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-427)\nA/1/367\n\"=\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-367)\nA/1/747\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-747)\nA/1/1220\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1220)\nA/1/1782\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1782)\nA/1/2324\n\"=\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2324)\nA/1/2876\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2876)\nA/1/3122\n\"-\" for dashes within journal citations?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3122)\nA/1/774\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-774)\nA/1/3838\n\" @\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3838)\nA/1/1415\n\"),\", sometimes with more leading punctuation, in scientific contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1415)\nA/1/3877\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3877)\nCluster #124\nA/1/1147\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1147)\nA/1/2930\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2930)\nA/1/3857\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3857)\nCluster #127\nA/0/359\nThis feature fires when it sees the punctuation mark comma \",\" \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-359)\nA/1/178\n\",\" in mechanical engineering?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-178)\nA/1/713\n\",\" after \"However\" or \"by contrast\" type phrases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-713)\nA/1/1204\nThis feature attends to commas used as enumerators in scientific texts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1204)\nA/1/1374\n\",\" in lists of chemicals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1374)\nA/1/1565\n\"],\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1565)\nA/1/1801\n\",\" between demographic factors\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1801)\nA/1/2284\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2284)\nA/1/2835\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2835)\nA/1/3482\n\";\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3482)\nA/1/3921\n\";\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3921)\nCluster #37\nA/0/289\nThis feature attends to parentheses and other bracket notations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-289)\nA/1/1785\nThis feature attends to closing parentheses.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1785)\nA/1/1885\n\")\" in biomedical studies\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1885)\nA/1/3672\n. The \"]\" at the end of each reference code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3672)\nA/1/715\n\",\" after an adverb indicating a conclusion from something previous, e.g. \"Finally\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-715)\nA/0/231\nThe feature attends to the word \"in\" followed by a comma and appearing at the beginning of a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-231)\nCluster #142\nA/0/430\nThis feature fires when a comma is present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-430)\nA/1/1284\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1284)\nA/1/1479\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1479)\nA/1/2409\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2409)\nA/1/3342\n\",\" in lists, mostly of short nouns/phrases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3342)\nA/1/3411\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3411)\nCluster #10\nA/0/128\nThis feature appears to attend to proper nouns and named entities in text, especially names of people, organizations, locations, and creative works.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-128)\nA/1/626\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-626)\nA/1/1366\nSports, esp. Title Case names of competitions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1366)\nA/1/1860\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1860)\nA/1/2731\n\" the\" in sports (football?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2731)\nA/1/2784\nAmerican sports concepts: offensive/defensive lines and defensive/tackle, right-handed bat, three-pointer, starter, drills\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2784)\nA/1/3316\nProfessional sports\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3316)\nA/1/3336\n1- to 2-digit numbers in sports contexts, esp. scores\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3336)\nA/1/3856\n\" with\" in sports contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3856)\nA/1/993\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-993)\nA/1/1814\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1814)\nA/1/757\nCommas after conjunctions at starts of sentences or connecting clauses\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-757)\nCluster #143\nA/1/586\nCommas in discussions of military battles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-586)\nA/1/658\nThis feature fires when there is a comma used in the middle or end of a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-658)\nA/1/1327\nThis feature appears to fire when there is a comma followed by a word starting with a vowel.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1327)\nA/1/1389\nComma after an \"if\" clause\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1389)\nA/1/3754\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3754)\nA/1/3160\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3160)\nCluster #6\nA/0/75\nThis feature fires when it encounters the word \"and\" or the word \"or\" being used as conjunctions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-75)\nA/0/214\nThe feature fires when it sees the word 'and'.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-214)\nA/0/403\nThe feature fires for sentences that have a comma separating two phrases, clauses or lists, or the words \"or\", \"and\", \"but\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-403)\nA/1/33\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-33)\nA/1/56\nCode: \" and\"/\" or\" in comments about data\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-56)\nA/1/87\n\" and\" after country names and other geographical locations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-87)\nA/1/92\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-92)\nA/1/310\n\"and\" conjoining family members\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-310)\nA/1/324\n\" and\" after a full (first+last) name\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-324)\nA/1/337\nThis feature attends to coordinating conjunctions such as \"and\" and \"or\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-337)\nA/1/391\n\" and\" in U.S. history or geography\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-391)\nA/1/406\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-406)\nA/1/524\n\" or\" joining number words or similar, e.g. \"one or two\" or \"one or more\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-524)\nA/1/603\n\"and\" in chemistry text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-603)\nA/1/783\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-783)\nA/1/802\n\" or\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-802)\nA/1/822\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-822)\nA/1/843\n\" and\" in recipes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-843)\nA/1/1000\n\" and\" and other conjunctions in computing text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1000)\nA/1/1074\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1074)\nA/1/1137\n\" and\" shortly after EOT\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1137)\nA/1/1211\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1211)\nA/1/1249\n\" and\" scientific\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1249)\nA/1/1274\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1274)\nA/1/1349\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1349)\nA/1/1386\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1386)\nA/1/1431\n\" or\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1431)\nA/1/1495\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1495)\nA/1/1851\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1851)\nA/1/1882\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1882)\nA/1/1934\n\" and\" after an emotion, predicting another emotion\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1934)\nA/1/1949\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1949)\nA/1/2058\n\" &\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2058)\nA/1/2117\n\"and\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2117)\nA/1/2208\n\" and\"/\" or\", after a number and time interval (e.g. \"six months\") and predicting another time interval\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2208)\nA/1/2251\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2251)\nA/1/2272\n\" and\" and conjunctions in demography or population surveys\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2272)\nA/1/2356\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2356)\nA/1/2419\n\" and\" after groups of people, predicting other groups of people\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2419)\nA/1/2428\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2428)\nA/1/2431\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2431)\nA/1/2520\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2520)\nA/1/2585\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2585)\nA/1/2597\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2597)\nA/1/2612\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2612)\nA/1/2623\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2623)\nA/1/2679\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2679)\nA/1/2747\n\"and\" in context of color descriptions of objects\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2747)\nA/1/2849\n\" or\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2849)\nA/1/2860\nThe feature fires when words like \"or\", \"and\", \"to\", or punctuation like \",\" appear.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2860)\nA/1/2884\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2884)\nA/1/3042\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3042)\nA/1/3106\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3106)\nA/1/3142\n'or' in 'and/or'\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3142)\nA/1/3196\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3196)\nA/1/3217\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3217)\nA/1/3220\n\" and\"/\" or\" after an abbreviation, predicting another abbreviation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3220)\nA/1/3232\n\" and\"/\" or\" largely in all-lowercase contexts like chat logs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3232)\nA/1/3244\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3244)\nA/1/3454\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3454)\nA/1/3486\n\" and\"/\" or\" after and predicting adjectives like \"economic\" or \"medical\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3486)\nA/1/3618\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3618)\nA/1/3671\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3671)\nA/1/3757\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3757)\nA/1/3790\nThis feature fires for sentences that contain the word \"and\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3790)\nA/1/3875\n\"and\" conjunction predicting a past-tense verb\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3875)\nA/1/3906\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3906)\nA/1/3983\n\" and\" and sometimes \" or\" in medical/anatomical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3983)\nA/1/2611\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2611)\nCluster #147\nA/0/310\nThis feature fires when there is a comma followed by a space.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-310)\nA/1/2367\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2367)\nA/1/3376\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3376)\nA/1/3595\nThis feature seems to attend to punctuation, specifically commas. It fires highest for commas and has a secondary activation for other types of punctuation like exclamation points and semicolons.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3595)\nA/1/3634\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3634)\nA/1/3995\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3995)\nA/1/3506\nCommas near beginning of quoted English text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3506)\nA/1/3966\nAdjective describing a famous person (occupation or nationality)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3966)\nCluster #146\nA/0/504\nThis feature seems to fire when there is a comma followed by the word \"and\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-504)\nA/1/814\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-814)\nA/1/1059\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1059)\nA/1/1116\n',' comma in archaic english\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1116)\nA/1/1988\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1988)\nA/1/2925\nFemale - she / her / herself ... ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2925)\nA/1/2390\n\";\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2390)\nA/1/1928\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1928)\nA/1/382\n\",\" in sports reporting, particularly after \"season\" / units of time\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-382)\nA/1/3897\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3897)\nCluster #105\nA/0/123\nThis feature fires when a name is followed by a comma.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-123)\nA/1/739\nComma after a work of art (performance, album, book)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-739)\nA/1/1081\n\",\" separating full names in lists\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1081)\nA/1/1238\nCommas in lists of countries\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1238)\nA/1/1631\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1631)\nA/1/1828\n\",\" after scientific institutions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1828)\nA/1/2007\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2007)\nA/1/2397\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2397)\nA/1/2964\nCommas in European history\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2964)\nA/1/3459\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3459)\nA/1/1365\nThe feature fires when it sees quotation marks with text followed by punctuation such as a comma or period.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1365)\nA/1/657\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-657)\nCluster #84\nA/1/331\n\",\" in legal contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-331)\nA/1/673\n\",\" after/between years\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-673)\nA/1/1864\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1864)\nCluster #80\nA/0/6\nThe feature fires when it sees punctuation like commas, semicolons and periods, especially when they follow a word that ends with a lowercase letter.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-6)\nA/1/1066\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1066)\nA/1/1152\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1152)\nA/1/3394\nThis feature fires when it sees a semicolon (;) followed by a parenthetical citation with a 4 digit date.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3394)\nA/1/3697\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3697)\nA/1/1692\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1692)\nA/1/2893\nCommas in thousands-place separators (patent numbers?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2893)\nCluster #57\nA/1/1406\n\".,\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1406)\nA/1/2622\n\",\" after physicist last names, predicting \" Phys\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2622)\nA/1/2767\nCommas in citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2767)\nCluster #73\nA/0/240\nThis feature attends to numbers from 1 to 99,999 preceded or followed by commas.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-240)\nA/1/32\n\",\" after small numbers in sequences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-32)\nA/1/2014\nCommas separating numbers (with spaces)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2014)\nA/1/2209\n\",\" in numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2209)\nCluster #83\nA/1/546\nCommas, often in book titles or preceding them to delimit the author\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-546)\nA/1/2690\n\"]\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2690)\nA/1/3653\n\")\" in legal documents\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3653)\nA/1/4051\nKorean: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4051)\nCluster #7\nA/0/48\nThis feature fires when it sees Greek text, specifically Ancient Greek.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-48)\nA/1/572\nGreek: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-572)\nA/1/634\nGreek: \"ι\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-634)\nA/1/1586\nKhmer, Amharic, (Greek?): ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1586)\nA/1/2599\nGreek: Capital letters?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2599)\nA/1/3453\nGreek: nu, chi\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3453)\nA/1/3831\nGreek: Vowel, especially omicron and upsilon\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3831)\nCluster #13\nA/0/66\nThis feature fires when it sees foreign characters such as accented characters, Cyrillic and Greek characters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-66)\nA/0/79\nThis feature seems to fire on mathematical fractions, degrees symbols, and other symbols commonly used in math and science writing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-79)\nA/0/80\nThis feature fires on Greek text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-80)\nA/0/101\nThis feature seems to fire for characters and scripts that are not Latin-based. Some examples where it has high activation include Georgian, Japanese, Chinese, Korean, Greek, Cyrillic, and Thai characters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-101)\nA/0/290\nThis feature attends to non-Latin alphabet languages such as Thai, Tamil, Georgian, Hindi, Japanese, Arabic, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-290)\nA/1/0\nJapanese, Chinese, other Asian scripts: Interwoven in English\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-0)\nA/1/78\nJapanese: \"\\xe9\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-78)\nA/1/281\nGreek?: \"\\\\\\xe1\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-281)\nA/1/422\nThai: Unicode prefix \"\\xe0\\xb8\", after vowel\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-422)\nA/1/426\nStart of Unicode character esp. predicting \\x95 in U+2215 HORIZONTAL BAR\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-426)\nA/1/432\n\"\\xc3\" in UTF-8 as latin1 mojibake, esp. of Russian\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-432)\nA/1/558\n\"\\\\\\xce\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-558)\nA/1/589\n\"\\\\\\xe0\\\\\\xa4\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-589)\nA/1/619\n\"\\\\\\xe1\\\\\\xbb\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-619)\nA/1/641\nJapanese: beginnings of characters?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-641)\nA/1/795\nArabic: Unicode start \"\\xd9\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-795)\nA/1/870\nJapanese: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-870)\nA/1/903\n\"\\\\\\xce\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-903)\nA/1/934\nJapanese: \"\\xe8\\xa6\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-934)\nA/1/1009\n\" \\\\\\xce\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1009)\nA/1/1016\nHebrew, Arabic, Farsi: Unicode block \"\\xd7\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1016)\nA/1/1096\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1096)\nA/1/1163\nTelugu?: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1163)\nA/1/1193\nJapanese: \"\\\\\\xe3\\\\\\x82\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1193)\nA/1/1232\nMath (Incomplete Unicode)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1232)\nA/1/1235\n\\xc2 in UTF-8 as latin1 mojibake of Korean?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1235)\nA/1/1345\n\"\\xc3\" in UTF-8 as latin1 mojibake of Arabic?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1345)\nA/1/1369\nBengali: Unicode sequence start\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1369)\nA/1/1499\n\"\\\\\\xc4\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1499)\nA/1/1724\nGeorgian and Korean?: \"\\xe1\\x83\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1724)\nA/1/1820\nSouth Asian scripts: Unicode start\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1820)\nA/1/1904\nJapanese: Katakana \"extend the previous vowel sound for longer\" partial-uniocde character\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1904)\nA/1/1940\nJapanese: beginnings of kanji\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1940)\nA/1/2012\n\"\\\\\\xcf\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2012)\nA/1/2152\nJapanese: Beginnings of kanji\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2152)\nA/1/2399\nJapanese, Chinese: \"\\xe6\\xa5\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2399)\nA/1/2453\nJapanese: \"\\xe9\\x9a\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2453)\nA/1/2465\nArmenian: \"\\xd5\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2465)\nA/1/2466\n\\xc2 as UTF-8 as latin1 mojibake of Chinese?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2466)\nA/1/2472\nVietnamese (Thai?): \"\\xe1\\xba\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2472)\nA/1/2556\nJapanese, Chinese, other Asian scripts: Unicode Start \\xe6\\xb0?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2556)\nA/1/2563\nThai: Unicode prefix \"\\xe0\\xb9\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2563)\nA/1/2608\n\"\\\\\\xce\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2608)\nA/1/2617\nChinese and Japanese: beginnings of characters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2617)\nA/1/2618\nJapanese: Hiragana unicode character byte-prefix (predicting the postfix)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2618)\nA/1/2636\n\\xc2 in UTF-8 as latin1 mojibake of Korean?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2636)\nA/1/2725\n\\xc2 in UTF-8 as latin1 mojibake of Chinese?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2725)\nA/1/2736\nJapanese and Chinese: \"\\xe5\\x88\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2736)\nA/1/2745\n\\xc2 in UTF-8 as latin1 mojibake of Chinese?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2745)\nA/1/2782\nJapanese: \"\\\\\\xe3\\\\\\x83\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2782)\nA/1/2798\n\\xc2 (Unicode starter of Unicode continuation) in UTF-8 as latin1 mojibake of Russian?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2798)\nA/1/2852\n\" \\\\\\xcf\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2852)\nA/1/2919\n\" \\\\\\xe0\\\\\\xa4\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2919)\nA/1/2957\nUnicode start codes for combining accents\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2957)\nA/1/3006\nEmoji? / mid-Unicode bytes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3006)\nA/1/3133\nHebrew: \"\\xd6\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3133)\nA/1/3150\n\"\\\\\\xc5\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3150)\nA/1/3345\nGujarati: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3345)\nA/1/3368\nJapanese and other Asian languages: \"\\xe9\\x80\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3368)\nA/1/3399\nArabic: Unicode start \"\\xd8\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3399)\nA/1/3468\nJapanese, Chinese, and Korean: \"\\\\\\xe7\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3468)\nA/1/3561\nThai: unicode prefix \"\\xe0\\xb8\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3561)\nA/1/3687\nGeorgian (Unicode Start)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3687)\nA/1/3692\nHebrew: Unicode Block \" \\xd7\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3692)\nA/1/3726\n\"\\xc2\" in UTF-8 as latin1 mojibake of Chinese?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3726)\nA/1/3748\n\" \\xcc\" in mojibake?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3748)\nA/1/3785\nJapanese: \"\\xe6\\x8a\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3785)\nA/1/3815\nChinese and Japanese: Beginnings of characters, vaguely poetic?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3815)\nA/1/3907\nArabic: Unicode start \" \\xd8\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3907)\nA/1/3908\n\" \\xc3\" in UTF-8 as latin1 mojibake of Korean? (beginnings of chat logs?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3908)\nA/1/3967\nJapanese: \"\\xe7\\x9b\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3967)\nA/1/4004\nJapanese: \"\\xe8\\xaa\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4004)\nA/1/4050\nGreek: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4050)\nA/1/522\nTamil: Unicode start\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-522)\nA/1/1280\nGreek: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1280)\nA/1/3119\nCzech, Lithuanian, Tamil: Unicode code point start\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3119)\nCluster #17\nA/0/435\nThe feature fires when it sees text written in South Indian languages such as Tamil, Telugu, Malayalam, and Kannada.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-435)\nA/1/416\nHebrew: Unicode end\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-416)\nA/1/466\nSinhala: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-466)\nA/1/480\nArmenian: Unicode end\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-480)\nA/1/580\nGeorgian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-580)\nA/1/994\nGeorgian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-994)\nA/1/1287\nBengali: Unicode end\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1287)\nA/1/1900\nPunjabi: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1900)\nA/1/2499\nTamil: Unicode end\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2499)\nA/1/2586\nHindi: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2586)\nA/1/2629\nBengali, Thai, Tamil, Lao?: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2629)\nA/1/2790\nDevanagari (Hindi?): ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2790)\nA/1/2796\nDevanagari, Tamil, Telugu, Malayalam, and Thai: Unicode end\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2796)\nCluster #4\nA/0/17\nThe feature fires on non-ascii and special characters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-17)\nA/1/873\nUTF-8 as latin1 mojibake of Chinese?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-873)\nA/1/980\nEuro sign etc. in mojibake: UTF-8 as latin1 of quotes, U+201x\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-980)\nA/1/1049\nUTF-8 as latin1 mojibake of Chinese?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1049)\nA/1/1411\n\"ä\" U+00E4 in UTF-8 as latin1 mojibake, esp. of Chinese\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1411)\nA/1/1517\n\"\\x91\" in UTF-8 as latin1 mojibake, esp. of Russian\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1517)\nA/1/1635\n\"\\\\\\x90\" UTF-8 as latin1 mojibake of Russian?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1635)\nA/1/2387\nUTF-8 as latin1 mojibake of Korean?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2387)\nA/1/3041\nParticularly nasty mojibake (latin1, Windows-1252, Mac Roman; possibly iterated multiple times)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3041)\nA/1/3440\n\"Ã\" U+00C3 in two-character mojibake, broadly accented characters in European languages\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3440)\nA/1/3914\nUTF-8 as latin1 mojibake of Chinese?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3914)\nA/1/3961\nSome Cyrillic script? Unicode starters in UTF-8 as latin1 mojibake, extra cursed by normalization\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3961)\nA/1/1691\nKorean: \" \\\\\\xed\"/\"\\\\\\xed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1691)\nCluster #15\nA/0/319\nThe feature attends to Korean text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-319)\nA/1/1184\nKorean: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1184)\nA/1/2323\nArabic, Korean, Chinese, Japanese, Russian, Norwegian, et al.: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2323)\nA/1/2955\nKorean: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2955)\nCluster #18\nA/0/459\nThe feature appears to activate for text written in right-to-left languages such as Arabic, Persian, and Hebrew.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-459)\nA/1/1271\nArabic script: Whitespace\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1271)\nA/1/1466\nArabic script (Urdu?): ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1466)\nA/1/1706\nThai: (Unicode Sequence Complete)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1706)\nA/1/3134\nArabic script: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3134)\nA/1/3450\nArabic script: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450)\nCluster #20\nA/1/207\nTurkic family?: languages in Cyrillic script\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-207)\nA/1/1197\nRussian: \" \\u0441\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1197)\nA/1/1858\nRussian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1858)\nA/1/2363\nRussian: in (conjunction and verbal prefix)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2363)\nA/1/2459\nRussian: word-ends?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2459)\nA/1/766\nRussian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-766)\nCluster #14\nA/0/98\nThis feature attends to Russian words related to action verbs, especially those ending in -ть.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-98)\nA/0/413\nThis feature attends to Russian grammatical features like verb conjugations and declensions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-413)\nA/1/3181\nRussian: Prefix in verbs / nouns \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3181)\nA/1/3061\nRussian: Beginnings of quotes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3061)\nA/1/2664\nRussian: \" \\u0443\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2664)\nCluster #19\nA/1/24\nRussian: \"и\" ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-24)\nA/1/1473\nRussian: \" Т\" / \" т\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1473)\nA/1/1964\nRussian: \" \\u0437\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1964)\nA/1/2557\nRussian?: The common \"-ция\" suffix indicates that the word is a noun\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2557)\nA/1/3073\nRussian: л and к characters?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3073)\nA/1/3415\nRussian: Mostly letter \"е\" \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3415)\nA/1/3553\nRussian: \" \\u0434\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3553)\nA/1/4095\nRussian: \" п\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4095)\nA/0/130\nThis feature fires when it sees contractions with \"I'm\" or \"I am\", especially at the beginning of sentences or quotes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-130)\nA/1/225\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-225)\nA/1/1002\n\" am\" after \"I\", in context of subjective experience \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1002)\nA/1/983\n\" im\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-983)\nCluster #31\nA/0/55\nThis feature fires when the words \"was\" or \"were\" have a high activation value.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-55)\nA/0/164\nThis feature fires when the token \"is\" is present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-164)\nA/0/272\nThis feature fires when the phrase \"be ___\" (with some variation) is present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-272)\nA/1/4\n\" was\", esp. after \"he\"/\"she\"/a name, predicting an adjective\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4)\nA/1/111\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-111)\nA/1/253\n\" be\" after auxiliary verbs like \"must\"/\"may\", sometimes also with \"not\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-253)\nA/1/517\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-517)\nA/1/797\n\" be\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-797)\nA/1/885\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-885)\nA/1/1095\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1095)\nA/1/1378\nThe feature fires when the third-person singular present-tense \"is\" follows a noun phrase with an embedded clause, often signaling that a fact or argument is being stated.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1378)\nA/1/1521\n\" been\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1521)\nA/1/1750\n\" become\"/\" becoming\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1750)\nA/1/2049\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2049)\nA/1/2076\nThis feature fires when a verb in the past tense follows \"was\" or \"were\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2076)\nA/1/2131\n\" was\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2131)\nA/1/2188\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2188)\nA/1/2325\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2325)\nA/1/2481\n\"al\" - logit nonsense\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2481)\nA/1/2755\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2755)\nA/1/3015\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3015)\nA/1/3063\n\" be\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3063)\nA/1/3161\n\" are\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3161)\nA/1/3535\n\" is\" / other copulae, esp after <EOT>\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3535)\nA/1/3558\n\" be\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3558)\nA/1/3910\n\" was\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3910)\nA/1/3982\n\" being\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3982)\nA/1/2560\n\" remain\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2560)\nCluster #32\nA/0/295\nThis feature fires when the token \"was\" appears in the text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-295)\nA/1/1001\nCopulas (mostly \" is\") in technical text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1001)\nA/1/1146\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1146)\nA/1/1426\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1426)\nA/1/1463\n\" be\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1463)\nA/1/1648\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1648)\nA/1/1787\n\" are\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1787)\nA/1/2307\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2307)\nA/1/2341\n\" been\" in scientific manuscripts (shown, demonstrated, proposed)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2341)\nA/1/2539\n\" was\"/\" were\" in biomedical studies\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2539)\nA/1/2542\n\" was\"/\" were\" in medical contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2542)\nA/1/2844\n\" be\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2844)\nA/1/3339\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3339)\nA/1/1223\n\"'m\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1223)\nA/1/138\n',\"' or \",'\" concluding what somebody said\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-138)\nA/1/2561\n\".\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2561)\nA/1/2591\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2591)\nA/0/334\nThe feature fires for the period (.) character followed by quotation marks (\").\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-334)\nCluster #115\nA/0/47\nThe feature detects the use of quotation marks as it activates whenever it encounters paired single or double quotation marks.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-47)\nA/1/1203\n\".\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1203)\nA/1/2346\n\".\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2346)\nA/1/2816\nClosing punctuation+quote esp in series of quotes with no unquoted text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2816)\nA/1/3421\n\".\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3421)\nA/1/2001\nThis feature fires when sentences end in question marks or exclamation points.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2001)\nA/1/311\n\".”\" (period + smart close quote)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-311)\nCluster #114\nA/0/200\nThe feature seems to fire the most when a sentence ends with a period.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-200)\nA/1/170\n\".\" in sports discussion\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-170)\nA/1/1497\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1497)\nA/1/2230\n\".'\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2230)\nA/1/3932\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3932)\nA/1/3968\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3968)\nCluster #0\nA/0/25\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-25)\nA/0/89\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-89)\nA/0/120\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-120)\nA/0/124\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-124)\nA/0/208\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-208)\nA/0/299\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-299)\nA/0/309\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-309)\nA/0/336\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-336)\nA/0/467\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-467)\nA/1/31\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-31)\nA/1/37\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-37)\nA/1/57\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-57)\nA/1/79\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-79)\nA/1/82\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-82)\nA/1/95\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-95)\nA/1/122\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-122)\nA/1/131\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-131)\nA/1/135\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-135)\nA/1/168\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-168)\nA/1/188\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-188)\nA/1/190\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-190)\nA/1/191\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-191)\nA/1/214\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-214)\nA/1/217\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-217)\nA/1/223\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-223)\nA/1/227\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-227)\nA/1/238\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-238)\nA/1/240\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-240)\nA/1/242\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-242)\nA/1/250\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-250)\nA/1/260\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-260)\nA/1/266\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-266)\nA/1/286\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-286)\nA/1/288\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-288)\nA/1/298\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-298)\nA/1/299\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-299)\nA/1/309\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-309)\nA/1/312\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-312)\nA/1/318\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-318)\nA/1/325\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-325)\nA/1/354\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-354)\nA/1/361\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-361)\nA/1/368\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-368)\nA/1/383\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-383)\nA/1/387\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-387)\nA/1/401\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-401)\nA/1/414\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-414)\nA/1/428\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-428)\nA/1/433\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-433)\nA/1/440\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-440)\nA/1/446\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-446)\nA/1/448\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-448)\nA/1/459\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-459)\nA/1/468\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-468)\nA/1/490\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-490)\nA/1/504\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-504)\nA/1/518\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-518)\nA/1/562\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-562)\nA/1/568\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-568)\nA/1/574\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-574)\nA/1/581\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-581)\nA/1/602\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-602)\nA/1/637\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-637)\nA/1/659\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-659)\nA/1/670\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-670)\nA/1/695\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-695)\nA/1/710\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-710)\nA/1/719\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-719)\nA/1/742\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-742)\nA/1/779\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-779)\nA/1/793\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-793)\nA/1/801\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-801)\nA/1/886\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-886)\nA/1/892\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-892)\nA/1/926\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-926)\nA/1/950\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-950)\nA/1/966\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-966)\nA/1/977\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-977)\nA/1/987\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-987)\nA/1/992\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-992)\nA/1/1006\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1006)\nA/1/1011\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1011)\nA/1/1018\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1018)\nA/1/1048\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1048)\nA/1/1058\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1058)\nA/1/1071\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1071)\nA/1/1072\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1072)\nA/1/1097\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1097)\nA/1/1109\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1109)\nA/1/1118\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1118)\nA/1/1121\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1121)\nA/1/1151\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1151)\nA/1/1158\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1158)\nA/1/1161\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1161)\nA/1/1170\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1170)\nA/1/1217\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1217)\nA/1/1221\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1221)\nA/1/1228\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1228)\nA/1/1247\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1247)\nA/1/1252\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1252)\nA/1/1257\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1257)\nA/1/1265\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1265)\nA/1/1267\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1267)\nA/1/1289\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1289)\nA/1/1290\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1290)\nA/1/1294\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1294)\nA/1/1324\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1324)\nA/1/1331\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1331)\nA/1/1336\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1336)\nA/1/1341\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1341)\nA/1/1356\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1356)\nA/1/1370\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1370)\nA/1/1393\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1393)\nA/1/1402\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1402)\nA/1/1407\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1407)\nA/1/1421\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1421)\nA/1/1427\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1427)\nA/1/1468\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1468)\nA/1/1486\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1486)\nA/1/1492\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1492)\nA/1/1504\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1504)\nA/1/1516\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1516)\nA/1/1523\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1523)\nA/1/1525\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1525)\nA/1/1528\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1528)\nA/1/1530\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1530)\nA/1/1540\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1540)\nA/1/1569\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1569)\nA/1/1574\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1574)\nA/1/1585\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1585)\nA/1/1590\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1590)\nA/1/1594\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1594)\nA/1/1617\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1617)\nA/1/1646\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1646)\nA/1/1684\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1684)\nA/1/1702\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1702)\nA/1/1705\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1705)\nA/1/1710\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1710)\nA/1/1711\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1711)\nA/1/1715\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1715)\nA/1/1734\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1734)\nA/1/1744\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1744)\nA/1/1748\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1748)\nA/1/1756\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1756)\nA/1/1765\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1765)\nA/1/1767\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1767)\nA/1/1769\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1769)\nA/1/1792\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1792)\nA/1/1816\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1816)\nA/1/1817\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1817)\nA/1/1819\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1819)\nA/1/1831\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1831)\nA/1/1839\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1839)\nA/1/1854\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1854)\nA/1/1861\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1861)\nA/1/1869\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1869)\nA/1/1873\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1873)\nA/1/1884\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1884)\nA/1/1888\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1888)\nA/1/1897\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1897)\nA/1/1899\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1899)\nA/1/1946\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1946)\nA/1/1963\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1963)\nA/1/1966\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1966)\nA/1/1971\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1971)\nA/1/1976\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1976)\nA/1/1989\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1989)\nA/1/1994\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1994)\nA/1/2004\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2004)\nA/1/2010\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2010)\nA/1/2029\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2029)\nA/1/2043\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2043)\nA/1/2050\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2050)\nA/1/2080\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2080)\nA/1/2086\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2086)\nA/1/2093\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2093)\nA/1/2097\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2097)\nA/1/2110\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2110)\nA/1/2120\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2120)\nA/1/2128\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2128)\nA/1/2138\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2138)\nA/1/2218\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2218)\nA/1/2220\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2220)\nA/1/2223\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2223)\nA/1/2238\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2238)\nA/1/2246\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2246)\nA/1/2247\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2247)\nA/1/2257\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2257)\nA/1/2297\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2297)\nA/1/2305\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2305)\nA/1/2333\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2333)\nA/1/2355\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2355)\nA/1/2365\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2365)\nA/1/2388\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2388)\nA/1/2392\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2392)\nA/1/2408\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2408)\nA/1/2437\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2437)\nA/1/2441\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2441)\nA/1/2447\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2447)\nA/1/2470\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2470)\nA/1/2475\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2475)\nA/1/2480\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2480)\nA/1/2490\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2490)\nA/1/2511\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2511)\nA/1/2521\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2521)\nA/1/2537\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2537)\nA/1/2541\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2541)\nA/1/2544\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2544)\nA/1/2551\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2551)\nA/1/2564\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2564)\nA/1/2575\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2575)\nA/1/2581\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2581)\nA/1/2583\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2583)\nA/1/2619\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2619)\nA/1/2643\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2643)\nA/1/2648\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2648)\nA/1/2682\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2682)\nA/1/2685\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2685)\nA/1/2692\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2692)\nA/1/2714\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2714)\nA/1/2716\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2716)\nA/1/2717\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2717)\nA/1/2726\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2726)\nA/1/2733\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2733)\nA/1/2749\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2749)\nA/1/2771\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2771)\nA/1/2775\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2775)\nA/1/2833\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2833)\nA/1/2858\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2858)\nA/1/2869\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2869)\nA/1/2870\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2870)\nA/1/2891\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2891)\nA/1/2895\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2895)\nA/1/2896\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2896)\nA/1/2912\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2912)\nA/1/2923\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2923)\nA/1/2924\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2924)\nA/1/2935\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2935)\nA/1/2936\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2936)\nA/1/2938\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2938)\nA/1/2961\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2961)\nA/1/2962\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2962)\nA/1/2966\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2966)\nA/1/2980\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2980)\nA/1/2987\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2987)\nA/1/2991\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2991)\nA/1/3025\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3025)\nA/1/3051\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3051)\nA/1/3069\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3069)\nA/1/3079\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3079)\nA/1/3086\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3086)\nA/1/3092\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3092)\nA/1/3101\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3101)\nA/1/3103\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3103)\nA/1/3113\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3113)\nA/1/3117\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3117)\nA/1/3127\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3127)\nA/1/3148\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3148)\nA/1/3153\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3153)\nA/1/3168\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3168)\nA/1/3170\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3170)\nA/1/3171\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3171)\nA/1/3172\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3172)\nA/1/3187\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3187)\nA/1/3202\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3202)\nA/1/3221\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3221)\nA/1/3239\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3239)\nA/1/3251\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3251)\nA/1/3257\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3257)\nA/1/3266\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3266)\nA/1/3268\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3268)\nA/1/3273\nThis feature fires for words related to light, specifically luminescence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3273)\nA/1/3277\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3277)\nA/1/3285\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3285)\nA/1/3296\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3296)\nA/1/3300\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3300)\nA/1/3302\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3302)\nA/1/3312\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3312)\nA/1/3330\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3330)\nA/1/3360\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3360)\nA/1/3372\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3372)\nA/1/3377\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3377)\nA/1/3390\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3390)\nA/1/3391\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3391)\nA/1/3395\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3395)\nA/1/3398\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3398)\nA/1/3409\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3409)\nA/1/3431\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3431)\nA/1/3436\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3436)\nA/1/3473\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3473)\nA/1/3478\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3478)\nA/1/3480\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3480)\nA/1/3497\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3497)\nA/1/3531\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3531)\nA/1/3538\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3538)\nA/1/3541\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3541)\nA/1/3544\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3544)\nA/1/3566\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3566)\nA/1/3571\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3571)\nA/1/3575\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3575)\nA/1/3586\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3586)\nA/1/3597\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3597)\nA/1/3633\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3633)\nA/1/3640\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3640)\nA/1/3645\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3645)\nA/1/3654\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3654)\nA/1/3661\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3661)\nA/1/3662\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3662)\nA/1/3686\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3686)\nA/1/3695\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3695)\nA/1/3707\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3707)\nA/1/3714\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3714)\nA/1/3719\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3719)\nA/1/3722\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3722)\nA/1/3724\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3724)\nA/1/3727\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3727)\nA/1/3731\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3731)\nA/1/3735\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3735)\nA/1/3744\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3744)\nA/1/3764\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3764)\nA/1/3770\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3770)\nA/1/3773\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3773)\nA/1/3775\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3775)\nA/1/3781\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3781)\nA/1/3786\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3786)\nA/1/3795\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3795)\nA/1/3824\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3824)\nA/1/3834\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3834)\nA/1/3862\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3862)\nA/1/3867\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3867)\nA/1/3896\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3896)\nA/1/3930\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3930)\nA/1/3935\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3935)\nA/1/3938\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3938)\nA/1/3943\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3943)\nA/1/3946\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3946)\nA/1/3951\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3951)\nA/1/3959\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3959)\nA/1/3971\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3971)\nA/1/3996\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3996)\nA/1/4002\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4002)\nA/1/4009\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4009)\nA/1/4015\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4015)\nA/1/4017\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4017)\nA/1/4021\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4021)\nA/1/4029\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4029)\nA/1/4039\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4039)\nA/1/4057\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4057)\nCluster #104\nA/0/15\nThis feature fires when the current token is followed by a period.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-15)\nA/1/248\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-248)\nA/1/264\n\".\" ending sentences in historical/chronological contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-264)\nA/1/415\nPeriod (in media review?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-415)\nA/1/537\n\".\" ending a sentence in a news report about politics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-537)\nA/1/614\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-614)\nA/1/913\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-913)\nA/1/1514\nPeriods and other sentence-concluding punctuation, esp. in video games?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1514)\nA/1/1722\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1722)\nA/1/1755\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1755)\nA/1/1809\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1809)\nA/1/2013\nPeriods, esp. shortly after EOT\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2013)\nA/1/2294\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2294)\nA/1/2411\n\".\" ending sentences in business/finance news contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2411)\nA/1/2495\n\").\" as end of English language parenthetical\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2495)\nA/1/3294\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3294)\nA/1/3515\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3515)\nA/1/3636\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3636)\nA/1/3958\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3958)\nCluster #103\nA/0/135\nThe feature fires when it sees punctuation marks and symbols like periods, parentheses, brackets, dollar signs and mathematical symbols at the end of tokens.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-135)\nA/0/498\nThis feature fires when it sees punctuation at the end of a sentence or phrase, particularly periods \".\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-498)\nA/1/513\n\".\" ending a sentence about medical care or formal medical reports\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-513)\nA/1/531\n\".\" ending a sentence in a scientific paper, mostly materials science\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-531)\nA/1/532\n\".\" ending a sentence in a journal paper or formal publication about disease or immunology or similar\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-532)\nA/1/831\nPeriods at the end of sentences in papers on radio astronomy\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-831)\nA/1/1357\n\"]\\\\\\\\].\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1357)\nA/1/2368\nEnd parenthetical + new sentence in academic param\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2368)\nA/1/3924\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3924)\nA/1/4024\nThis feature fires when a sentence ends with a period.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4024)\nA/1/796\n\" .\" used in repeated \" . . .\" style contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-796)\nA/1/3810\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3810)\nCluster #76\nA/0/424\nThis feature fires when it sees a sentence ending in a question mark.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-424)\nA/1/587\n\"!\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-587)\nA/1/2770\n\"?\" in news?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2770)\nA/0/285\nThis feature attends to sentences that end with a period followed by whitespace.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-285)\nA/1/621\n\"?\" in dialogue?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-621)\nA/1/3668\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3668)\nCluster #67\nA/1/979\n\" ;)\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-979)\nA/1/2062\n\" :)\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2062)\nA/1/2855\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2855)\nCluster #72\nA/0/46\nThis feature seems to fire when it sees words in chat logs or transcripts that users would type when asking for technical help or troubleshooting issues.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-46)\nA/1/277\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-277)\nA/1/1282\nUbuntu/linux conversations (irc chat?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1282)\nA/1/1902\n\" the\" in installations, operating systems?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1902)\nA/1/4079\nUbuntu/Linux/GNOME terms, esp. lowercase in chat\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4079)\nA/1/2641\n(IRC?) usernames being mentioned/replied to chat logs, after > and before :\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2641)\nA/1/2035\nPunctuation in paths or package names near \"ubuntu\"/\"linux\"/etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2035)\nCluster #71\nA/0/151\nThis feature fires when it sees <> and : style chat messages, IRC style messages, and other messaging styles like that.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-151)\nA/1/149\n\":\" in \">\"-based chat logs?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-149)\nA/1/2616\n\">\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2616)\nCluster #82\nA/0/474\nThis feature fires for instances of a single space in the token sequence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-474)\nA/1/1117\nThe first space in a double-space between sentences, in a legal context\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1117)\nA/1/3053\nNewline+spaces near Old/poetic English?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3053)\nA/1/3720\n\" \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3720)\nA/1/3899\nLine breaks followed by many spaces in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3899)\nA/1/4023\n\" \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4023)\nA/1/50\n\"...\" and some other long English punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-50)\nCluster #101\nA/0/437\nThis feature fires when it sees line breaks (like em-dashes, hyphens, vertical bars) used as separators, especially between sentences or clauses.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-437)\nA/1/185\nClosing parenthesis on years\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-185)\nA/1/598\n\"\\u201d\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-598)\nA/1/2548\n\"\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2548)\nA/1/3559\n\"space em-dash\" (em dashes typeset open) in English prose\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3559)\nA/1/3592\n\")\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3592)\nA/1/2759\nThe feature attends to punctuation around proper nouns.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2759)\nCluster #66\nA/0/367\nThis feature fires when it sees structured references (e.g., titles, sections, footnotes, page numbers) that are commonly seen in academic writing and publications.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-367)\nA/1/960\nPunctuation near \"_\"-surrounded numbers/variables\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-960)\nA/1/1916\nContinue chains of commas (and maybe related punctuation)?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1916)\nCluster #62\nA/0/11\nThis feature seems to fire on numerical decimal points, especially those with 2-3 digits before and after the decimal point.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-11)\nA/1/116\nLegal text \".\" and \").\" ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-116)\nA/1/498\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-498)\nA/1/569\nDecimal point in a number\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-569)\nA/1/1114\n\".\" as a decimal separator\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1114)\nA/1/1236\n\" .\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1236)\nA/1/1286\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1286)\nA/1/1821\n\"+\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1821)\nA/1/2106\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2106)\nA/1/2684\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2684)\nA/1/2768\n\".\" among long strings of digits, often ones with multiple \".\"s\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2768)\nA/1/2889\n\".\" between numbers near lots of spaces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2889)\nA/1/3082\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3082)\nA/1/3609\n\".\" interspersed between more than two numbers (version numbers, IP addresses; but not decimal numbers)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3609)\nA/1/2130\n\" ?\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2130)\nCluster #100\nA/0/507\nThe feature fires on the colon (:) symbol.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-507)\nA/1/694\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-694)\nA/1/959\n\" -\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-959)\nA/1/1004\n: separator in times\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1004)\nA/1/1914\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1914)\nA/1/2229\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2229)\nA/1/2347\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2347)\nA/1/2713\n: to indicate someone is speaking in a script\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2713)\nA/1/4091\n\":\" in citations of sources\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4091)\nCluster #97\nA/1/408\n\":\" after \"\\n\\nA\" (in Q: and A: pairs?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-408)\nA/1/2144\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2144)\nA/1/2186\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2186)\nA/1/2908\n\" !\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2908)\nA/0/402\nThe feature seems to fire when it encounters punctuation such as colons, parentheses, brackets, slashes, pipes, etc. Especially when they are at the end of words or lines.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-402)\nA/1/96\n\":\" esp. after \"abstract\" (in some paper metadata?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-96)\nA/1/2846\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2846)\nA/1/3751\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3751)\nA/1/3669\nPunctuation or non-content links in forum posts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3669)\nA/1/2817\nTime specification\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2817)\nA/1/1013\n\" \"/\" \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1013)\nA/1/1537\nLong spans of whitespace in poorly-formatted documents, often near decimal points\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1537)\nCluster #86\nA/0/426\nThis feature attends to formatted tabular data such as titles, tables, lists, tables of contents, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-426)\nA/1/1299\nSpaces near vertical bars (Markdown/ASCII tables?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1299)\nA/1/3826\nRuns of periods\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3826)\nA/1/4090\n\" |\" in Markdown tables?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4090)\nA/1/2885\nEmoji ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2885)\nCluster #85\nA/1/91\nASCII/Unicode box drawings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-91)\nA/1/972\nLong runs of spaces (in technical writing / math?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-972)\nA/1/1717\nMath\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1717)\nA/1/2139\n\"-->\" indicating the span of time between two listed timestamps (and occasionally captures hexadecimal base pairs separated by spaces?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2139)\nA/1/682\nPipe-separated text (esp. words)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-682)\nA/1/1852\nLaTeX \"bf\" and other macros after \"\\\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1852)\nCluster #45\nA/0/95\nThe feature attends to hexadecimal numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-95)\nA/1/818\nComma-separated numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-818)\nA/1/884\nDigits after \"0x\" in \"0x\" hex literals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-884)\nA/1/1399\n\" \"/\" \"/\" \"/\" \"/\" \"/\" \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1399)\nA/1/2213\n\",\" separating hexadecimal literals like \"0xf1\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2213)\nA/1/2430\n\" 0\" beginning \"0x\" hexadecimal literals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2430)\nA/1/3203\nHex numbers after \"0x\", usually in code defining ALLCAPS constants\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3203)\nA/1/2605\nThis feature fires on words that contain a dash \"-\", particularly scientific and biochemistry related terms.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2605)\nA/1/3942\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3942)\nA/1/1272\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1272)\nCluster #41\nA/0/114\nThe feature fires when it encounters a hyphen followed by lowercase letters like in chemical compound names and hyphenated phrases.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-114)\nA/1/99\nHyphen in Title-Case compounds\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-99)\nA/1/145\nHyphen after ALLCAPS abbreviation (SCPs, etc...)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-145)\nA/1/2113\n\"-\" connecting last names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2113)\nA/1/2327\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2327)\nA/1/2569\n\"/\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2569)\nCluster #30\nA/0/112\nThis feature fires on hyphens between numbers, especially when denoting age.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-112)\nA/1/12\nHyphen, esp. after \"Code § \" and a number, denoting a legal code section\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-12)\nA/1/906\n\"/\" for number/number in normal language text, often used for things like dates or simple fractions, but not in a math formula\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-906)\nA/1/1326\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1326)\nA/1/1678\n\"\\u2013\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1678)\nA/1/2115\n\"-\" in identifiers/numbers/phone-numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2115)\nA/1/2967\nSeperator in dates\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2967)\nA/1/3712\n^/^- in scientific notation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3712)\nCluster #40\nA/0/243\nThis feature seems to be attending to hyphens followed by another word or number.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-243)\nA/0/286\nThis feature fires on hyphenated words often indicating prefixes and suffixes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-286)\nA/1/246\n\"-\" in hyphenated constructs esp. after \"anti\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-246)\nA/1/333\n\"-\" after common compound-word-forming preposition: \"in\"/\"on\"/\"over\"/...\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-333)\nA/1/489\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-489)\nA/1/556\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-556)\nA/1/902\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-902)\nA/1/986\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-986)\nA/1/1435\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1435)\nA/1/1524\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1524)\nA/1/1887\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1887)\nA/1/3007\n\"-\" after \"well\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3007)\nA/1/3110\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3110)\nA/1/3245\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3245)\nA/1/3271\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3271)\nA/1/3423\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3423)\nA/1/3474\n\"-\" in \"non-\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3474)\nA/1/3643\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3643)\nA/1/3931\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3931)\nA/1/1043\n\"/\" separating language concepts like \"and/or\" or \"noun/othernoun\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1043)\nA/1/2438\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2438)\nA/0/276\nThis feature attends to punctuation like hyphens, dashes, slashes, parentheses, quotes, apostrophes, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-276)\nA/1/1115\nHyphen used to join words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1115)\nA/0/385\nThis feature seems to fire after underscore characters when they are being used to emphasize or highlight the words/phrases before and after them.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-385)\nA/1/3679\n\"_\" typically concluding italics for a title, e.g. in Markdown\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3679)\nCluster #29\nA/0/157\nThe feature fires when there are quotation marks or other types of quotes around words or phrases.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-157)\nA/1/19\nVarious European languages: Opening quotes?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-19)\nA/1/639\n\" \\\"\" quoting some legal text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-639)\nA/1/805\n\" \\u201c\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-805)\nA/1/927\n\" \\\"\" for opening an airquote or small-phrased-quote\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-927)\nA/1/1129\nOpening quotes, mostly non-ascii and stronger on double-quotes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1129)\nA/1/2819\n\"\\u2014\", emdash\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2819)\nA/1/3649\nOpening quotes, esp. \" '\" and esp. after a comma\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3649)\nA/1/4084\n\" _\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4084)\nCluster #28\nA/0/401\nThe feature fires when the token directly after a period is a quotation mark.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-401)\nA/1/1763\n\"\\u201c\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1763)\nA/1/2701\n\" \\\"\" for someone speaking to follow a sentence of their behavior or previous speaking, as in a story\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2701)\nA/1/3123\n\" \\\"\" to as part of rapid-fire back-and-forth conversation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3123)\nA/1/3728\n\" \\\"\"as part of longer back-and-forth conversation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3728)\nA/0/53\nThis feature fires strongly on quotation marks, especially double quotation marks “ “.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-53)\nA/1/194\n'\"' after two newlines\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-194)\nCluster #132\nA/0/4\nThis feature fires on line breaks and paragraphs especially in legal or technical documents.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-4)\nA/1/2235\nCode?: Double newlines in legal text or code?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2235)\nA/1/3350\nQuadruple newlines in legal text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3350)\nA/1/3767\nTriple newlines, often after EOT\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3767)\nA/1/3463\nNewlines in recipies\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3463)\nCluster #163\nA/0/354\nThis feature attends to line breaks (new lines) in text, particularly double line breaks.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-354)\nA/1/1693\nNewline before numbers in legal text?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1693)\nA/1/2853\nNewline in legal text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2853)\nA/1/3462\nNewline in legal judgment?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3462)\nCluster #130\nA/1/804\nCourt filings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-804)\nA/1/1198\nNewline+Form Feed in legal text?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1198)\nA/1/1582\n\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1582)\nCluster #133\nA/0/429\nThis feature fires for line breaks such as carriage returns and newlines.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-429)\nA/1/184\nCode: \"\\r\\n\\r\" (carriage-return, newline, carriage-return)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-184)\nA/1/257\nCarriage returns, usu. before newlines, in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-257)\nA/1/625\nCode: Carriage returns, usu. before newlines, esp. in C or similar languages\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-625)\nA/1/2443\nCarriage returns, usu. followed by newlines, close to EOT and/or in HTML\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2443)\nA/1/2934\nCarriage return\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2934)\nA/1/3947\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3947)\nA/1/4068\nCode: Carriage returns, usu. before newlines, in CSS\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4068)\nCluster #167\nA/0/100\nThis feature fires when there is a newline character \"\\n\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-100)\nA/0/192\nThe feature fires when a new paragraph starts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-192)\nA/0/372\nThis feature fires for line breaks / new lines.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-372)\nA/1/171\nNewlines after proper nouns?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-171)\nA/1/1598\nNewline mid-sentence in art/fiction/nature?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1598)\nA/1/1619\nNewline shortly after EOT\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1619)\nA/1/2092\nNewline in discussion of accounts/authentication\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2092)\nA/1/2259\nNewline in economics?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2259)\nA/1/2354\nNewline after/between quotes/dialogue\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2354)\nA/1/2555\nNewline in English prose (fiction? religion/philosophy?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2555)\nA/1/2615\nNewlines in website updates, payment/subscription management\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2615)\nA/1/2727\nNewlines in patents? (mechanical engineering?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2727)\nA/1/2734\nNewline mid-sentence esp. in religious text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2734)\nA/1/2760\nNewlines before level-3 Markdown/ATX headers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2760)\nA/1/2822\nNewlines after punctuation that's not just a period\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2822)\nA/1/3141\nNewline in news reports?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3141)\nA/1/3155\nVarious European languages: Newline?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3155)\nA/1/3236\nNewline after sentence-ending period\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3236)\nA/1/3499\nNewline in bibliography list?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3499)\nA/1/3779\nNewline after \"was\"/\"were\" mid-sentence\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3779)\nCluster #181\nA/0/106\nThis feature fires when there's a new line that follows an end punctuation mark (., !, ?, \", etc.).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-106)\nA/1/148\nNewline after another newline, predicting new sentence/item\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-148)\nA/1/2454\nNewlines after \"References\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2454)\nA/1/948\nNewline in chat logs? (IRC?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-948)\nCluster #183\nA/0/475\nThis feature fires on new line characters (\"\\n\").\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-475)\nA/1/400\nNewlines near numbers or number-heavy IDs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-400)\nA/1/545\nNewline in biology contexts and/or near repeated \"-\"/\"=\" as horizontal dividers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-545)\nA/1/652\nNewline after some curly-brace-based header and before \"-\"-based horizontal rules\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-652)\nA/1/824\nNewlines after math expressions in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-824)\nA/1/863\n\" \\n\" after ######\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-863)\nA/1/2182\nNewline in biology near section markers?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2182)\nA/1/2738\nNewline separating Wikipedia categories\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2738)\nCluster #184\nA/0/375\nThis feature fires for text containing numbers ending in question marks on new lines.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-375)\nA/1/74\nNewlines in wikitext tables?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-74)\nA/1/591\nNewline near \"=\"-based horizontal rules\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-591)\nA/1/1261\n\"\\n\\n\" after \"add your own caption\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1261)\nA/1/3222\nNewline in LaTeX esp. after beginning a math environment\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3222)\nA/1/2813\nBibTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2813)\nA/1/651\nSynthetic math: Newline before/after one-line numbers in synthetic math problems\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-651)\nA/1/3352\nShort sequences of asterisks?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3352)\nA/1/845\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-845)\nA/0/132\nThe feature attends to Wikipedia articles, firing most for category headings.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-132)\nA/1/2861\nWikipedia categories\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2861)\nA/1/919\nQuotes/newlines in .po/.pot files (https://www.drupal.org/community/contributor-guide/reference-information/localize-drupal-org/working-with-offline/po-and-pot-files)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-919)\nA/1/2124\nCategory + :\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2124)\nA/1/1494\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1494)\nA/1/275\nCode: \"-\" esp. in \"UTF-8\" and other Unicode encoding names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-275)\nCluster #172\nA/0/364\nThis feature fires on blocks of legal or license text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-364)\nA/1/72\n\",\" in several memorized license strings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-72)\nA/1/643\n\" FITNESS\" in code licenses (GPL?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-643)\nA/1/743\nWHETHER in a contract\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-743)\nA/1/803\n\"Competing interests\" phrases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-803)\nA/1/832\n\" under\" in software licenses\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-832)\nA/1/1437\nCopyright dates (and sometimes copyright organizations)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1437)\nA/1/1647\nAll rights reserved. (and some other source code licensing texts)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1647)\nA/1/1834\n\" *\" (and sometimes \"//\") for use in copyright license multiline c/c++ comments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1834)\nA/1/2105\n\" the\" in the GPL license?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2105)\nA/1/2135\nLists of 'Category' in movie metadata?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2135)\nA/1/2711\n\" COURT\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2711)\nA/1/2720\nBSD License\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2720)\nA/1/2985\n\" picked\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2985)\nA/1/3005\n\" respect\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3005)\nA/1/3037\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3037)\nA/1/3219\n\" License\" in the GPL\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3219)\nA/1/3349\nFunding acknowledgements\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3349)\nA/1/3691\nActivates on \"Copyright\", predicts the copyright symbol and dates\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3691)\nA/1/3845\n\"Click here for additional data file\" and supplemental data more generally\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3845)\nA/1/3029\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3029)\nA/1/1662\nCode: XML namespace references\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1662)\nA/1/3263\nCode: Error variables/handling in Go?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3263)\nA/1/1800\nDOIs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1800)\nA/1/2189\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2189)\nA/1/2514\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2514)\nA/1/2840\nNewline mid-sentence or near colons?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2840)\nCluster #54\nA/0/360\nThe feature fires on parentheses, especially at the end of lines and sentences.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-360)\nA/1/853\n1) Lists 2) by 3) number?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-853)\nA/1/1795\nNewline after closing parentheses in English prose\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1795)\nCluster #154\nA/0/172\nThis feature fires for code snippets and other text with programming language syntax, as well as indented outlines and other structured/formatted text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-172)\nA/1/403\nCode: Newline in HTML\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-403)\nA/1/716\nCode: Newline in CSS\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-716)\nA/1/973\nCode: Newline in shell commands / documentation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-973)\nA/1/2824\nTokens near/predicting \"Q\" in various contexts, esp. \"_\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2824)\nA/1/2917\nNewlines after \"Q:\" and maybe more generally technical troubleshooting\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2917)\nA/1/3121\nSecond newline after \"A:\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3121)\nA/1/4063\nNewline in code-adjacent English prose\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4063)\nCluster #153\nA/0/18\nThis feature fires for source code, specifically highlighting programming language syntax such as curly braces, indentation, comments, include statements, function definitions, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-18)\nA/1/559\nCode: Newlines in code comments?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-559)\nA/1/578\nCode: Newlines after/between lines beginning with \"#\" (comments or C/C++ preprocessor?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-578)\nA/1/1542\nCode: Newlines between/after imports, esp. Python\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1542)\nA/1/3044\nCode: Newline after closing delimiters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3044)\nA/1/3715\nCode: Newline in C/C++ especially after comment closing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3715)\nA/1/23\nCode: Newlines in or near C-style /* comments */\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-23)\nA/1/1375\nCode: Newline in C/C++/Go?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1375)\nA/1/2421\nCode: \"#\" in C/C++ preprocessor\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2421)\nCluster #52\nA/1/1303\nCode: '.' and '#' in C include statements\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1303)\nA/1/1490\nWeb file extensions in URLs or links\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1490)\nA/1/1741\n\"h\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1741)\nA/1/3812\n\"\\u23ce\\n\\t\"/\"\\u23ce\\n\\t\\t\"/\"\\u23ce\\n\\t\\t\\t\"/\"\\u23ce\\n\\t\\t\\t\\t\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3812)\nA/1/1302\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1302)\nA/1/2741\n\" {\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2741)\nCluster #160\nA/0/170\nThis feature fires on opening and closing punctuation such as brackets, braces, quotation marks, tags, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-170)\nA/0/390\nThis feature seems to fire on semicolons and curly braces, which are common in programming languages. So it is likely detecting source code.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-390)\nA/0/491\nThis feature attends to code blocks and style elements found in programming languages like indentation, brackets, semicolons, keywords, variables, and function names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-491)\nA/1/1008\nCode: Semicolons concluding statements, sometimes with preceding punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1008)\nA/1/1677\nCode: \" }\" concluding code blocks/functions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1677)\nA/1/1938\nThis feature is picking up on parentheses and curly braces that contain code or math expressions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1938)\nA/1/2037\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2037)\nA/1/2077\nCode: Closing delimiters and semicolons, predicting newlines and tab-based indentation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2077)\nA/1/2336\nCode: ends of HTML/XML tags\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2336)\nA/1/2414\nCode: Semicolons, esp. ending \"return\" statements\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2414)\nA/1/2562\n\">\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2562)\nA/1/3135\n\" {\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3135)\nA/1/4030\nCode: \";\"-based statement enders in code with and predicting \"\\r\\n\" (+ spaces)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4030)\nA/1/4088\n\" */\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4088)\nA/1/1020\n\" []\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1020)\nA/1/1611\nThis feature attends to the use of semicolons at the ends of lines of code.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1611)\nCluster #161\nA/0/30\nThis feature fires on opening and closing HTML tags.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-30)\nA/1/609\nThe feature attends to double quotes that signify the beginning or end of strings/keys in code.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-609)\nA/1/995\n\">\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-995)\nA/1/2995\nThis feature fires when it sees semicolons, especially at the end of lines.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2995)\nA/1/3016\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3016)\nA/1/1909\nCode: Newline followed by spaces in XML\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1909)\nCluster #162\nA/0/202\nThis feature seems to activate when it sees parentheses used to enclose parameters in function and method declarations and calls in programming languages.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-202)\nA/1/2823\nCode: \"()\" in function calls or definitions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2823)\nA/1/3461\nThis feature attends to brackets, parentheses and curly braces.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3461)\nA/1/1311\nCode: HTML closing tag before any start of text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1311)\nA/1/2477\n\"\\\">\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2477)\nCluster #158\nA/0/494\nThis feature seems to fire on HTML/XML tags and related code elements. It pays particular attention to opening and closing tags, as well as elements like paths, classes, properties, fields, methods, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-494)\nA/1/1910\nCode: HTML Template Format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1910)\nA/1/2909\nCode: end double quote (in HTML attributes)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2909)\nA/1/2965\n\" \\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2965)\nCluster #145\nA/1/872\nCode: uint when it's likely uint8/uint16/uint32/uint64\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-872)\nA/1/951\nCode: Go?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-951)\nA/1/1087\nCode: Function receiver types in Go\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1087)\nA/1/1529\nCode: punctuation, esp. ending in \",\", and spaces, esp. in Go\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1529)\nA/1/2303\nCode: Whitespace, esp. in Go structs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2303)\nA/1/2534\nCode: 8/16/32/64 in uintX_t types in C\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2534)\nA/1/463\nCode: Newline+spaces in doc-comments?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-463)\nA/1/1654\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1654)\nA/1/2969\n\"static\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2969)\nA/1/3235\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3235)\nCluster #178\nA/0/19\nThis feature fires when it sees punctuation like commas, semicolons, parentheses and brackets.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-19)\nA/1/229\nCode: \",\" in argument lists?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-229)\nA/1/472\nThe feature attends to punctuation such as commas and parenthesis.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-472)\nA/1/663\nThis feature fires when it encounters paired quotes (single or double) with comma-separated values inside.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-663)\nA/1/928\n\"\\\":\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-928)\nA/1/1877\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1877)\nCluster #187\nA/0/268\nThis feature attends to code structures like variable and function declarations, as well as code syntax like braces, semicolons, parentheses, arrows, and comparison operators.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-268)\nA/0/322\nThis feature detects code, as it fires on code syntax such as curly braces, semicolons, indentation, function definitions, comments, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-322)\nA/1/89\nCode: Newline and then spaces, esp after \"{\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-89)\nA/1/158\nCode: Comments: Runs of spaces?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-158)\nA/1/249\nCode: Newline+indentation in Python?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-249)\nA/1/395\nThis feature attends to token sequences indicative of function or method declarations and definitions in programming languages like C++, Ruby and Python.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-395)\nA/1/662\n\" if\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-662)\nA/1/852\nCode: equal/not-equal conditions in C/C++ code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-852)\nA/1/988\nCode: Newlines and spaces after curly braces (open or close) in C/C++/Java/similar\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-988)\nA/1/1878\nCode: \" =\" assigning a variable in a JavaScript/Java context (variable naming)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1878)\nA/1/2057\nCode: \" =\" assigning a variable in a C/C++ context\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2057)\nA/1/3023\n\" return\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3023)\nA/1/3031\n\" -\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3031)\nA/1/3299\nThis feature seems to activate most strongly when there are 4 or more spaces at the start of a line. This trend is most noticeable in examples 2, 3, 4, 5, 6, 7, 8, 12, 25, 32, 37, and 41. It appears that this feature is detecting indentation in code.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3299)\nA/1/3320\nCode: C++ streams\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3320)\nA/1/3607\nCode: Newline followed by spaces (C/C++/Java/C#?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3607)\nA/1/3334\nCode: identifiers close after tabs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3334)\nA/1/699\n\" /\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-699)\nCluster #188\nA/0/109\nThe feature fires for code in C languages based on whitespace, curly braces, semicolons, comments, function declarations, preprocessor directives, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-109)\nA/1/418\nClosing parenthesis or bracket, possibly with other punctuation, in technical contexts but not code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-418)\nA/1/1380\n\" //\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1380)\nA/1/1557\nCode: punctuation beginning comments like \"//\", \"/*\", or \"#\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1557)\nA/1/1937\nCode: C (or Java?) *'s in doc-comments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1937)\nA/1/3157\n\" #\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3157)\nA/1/3259\n\"####\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3259)\nA/1/675\nCode: Long runs of asterisks in comments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-675)\nA/1/3177\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3177)\nA/1/570\nCode: \"define\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-570)\nA/1/2873\nCode: \" *\" for pointers in C/C++ code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2873)\nA/1/2402\nCode: \"java\" in (file.java:line) errors/stack traces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2402)\nCluster #60\nA/0/166\nThis feature attends to URL links.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-166)\nA/1/1279\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1279)\nA/1/1320\nCode: \" at\" in Java stack traces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1320)\nA/1/1367\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1367)\nA/1/1503\n\".\" in domain names, esp. email addresses\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1503)\nA/1/1589\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1589)\nA/1/2255\n\".\" in domain names in URLs, mildly open-source-flavored\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2255)\nA/1/3281\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3281)\nA/1/3582\nCode: \".\" in Java import paths?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3582)\nA/1/3732\nCode: \".\" in C# import paths?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3732)\nA/1/4064\n\"/\" in URLs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4064)\nCluster #197\nA/0/118\nThe feature appears to fire on parentheses, arithmetic symbols, punctuation, whitespace and other programming/markup language syntax.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-118)\nA/0/139\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-139)\nA/0/298\nThis feature attends to text formatting characters such as slashes, quotes, brackets, braces, and backslashes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-298)\nA/1/764\nCode: punctuation near string manipulation and parsing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-764)\nA/1/767\nCode: variables/attributes predicting \"=\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-767)\nA/1/817\nCode: Assembly\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-817)\nA/1/996\n%-based Format String\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-996)\nA/1/1535\nWhitespace or English text near square-bracketed numbers/text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1535)\nA/1/1641\nMath entity prediction\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1641)\nA/1/1865\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1865)\nA/1/2990\nCode: Statistics programming languages? (some Pandas, some others)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2990)\nA/1/3434\nCode: Haskell\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3434)\nA/1/3610\nRegular expressions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3610)\nA/1/3314\nThis feature fires on lines, dashes, underlines, and other horizontal dividers in text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3314)\nA/1/1423\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1423)\nCluster #59\nA/0/406\nThis feature attends to sentence features that include object-oriented programming concepts like classes, methods, constructors, functions, variables, parameters, arguments, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-406)\nA/1/110\nCode: \".\" in long namespace/attribute member access chains (Java?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-110)\nA/1/481\n\" .\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-481)\nA/1/1436\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1436)\nA/1/1696\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1696)\nA/1/1738\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1738)\nA/1/1775\nCode: Function/method invocation in getter chains\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1775)\nA/1/2219\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2219)\nA/1/2814\nCode: \".\" in Java import paths?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2814)\nA/1/3022\nCode: \".\" as attribute access, accessing TitleCase members; likely Go\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3022)\nA/1/3318\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3318)\nA/1/3539\nCode: \".\" separating CapitalCamelCase identifiers (C#, Unity, etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3539)\nA/1/3615\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3615)\nA/1/3972\nCode: \".\" sometimes with preceding punctuation in varying contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3972)\nA/1/4087\nCode: `').` in jQuery invocation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4087)\nCluster #33\nA/0/326\nThis feature fires on code snippets, specifically identifier names, keywords, braces, semicolons, and other syntactical elements.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-326)\nA/1/1418\nTabs near uppercase letters (stock tickers? code?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1418)\nA/1/3830\n\"\\t\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3830)\nA/1/3872\n\"\\t\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3872)\nA/1/3988\nNewline after \"-\"-based horizontal rules\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3988)\nA/1/940\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-940)\nA/1/3192\n\"://\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3192)\nA/0/178\nThis feature fires on Twitter usernames in angled brackets (< and >).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-178)\nCluster #204\nA/0/414\nThe feature fires when it sees groups of characters in URLs that often represent file paths, folders, file types, domains and subdomain names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-414)\nA/1/1570\n\"@\", usually in email addresses\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1570)\nA/1/1979\nCode: '=\"' in HTML attributes, particularly <a href>\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1979)\nA/1/2683\n\"/\" and some other separators in URLs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2683)\nA/1/3338\nCode: hyphens in hyphenated code identifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3338)\nA/1/4043\nCode: \"/\" in import-paths\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4043)\nCluster #175\nA/1/152\nCode: \":\" in CSS specifically\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-152)\nA/1/1208\nCode: CSS alignment (predicting that a vertical/horizontal alignment follows)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1208)\nA/1/1430\nCode: hyphen in CSS\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1430)\nA/1/1935\nCode: \"-\" in font spec in CSS\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1935)\nA/1/2545\nCode: =\" when setting width/height in html/css\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2545)\nA/1/3577\n\"=\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3577)\nA/1/3509\nSubdomains or domains in URLs (after \"https://\" etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3509)\nCluster #216\nA/0/20\nThis feature attends to opening and closing HTML tags.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-20)\nA/1/121\nCode: HTML angle brackets opening/closing lists\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-121)\nA/1/290\nCode: HTML tag syntax that could/does enclose table-related tags\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-290)\nA/1/968\nCode: HTML/XML beginnings of close tags \"</\", sometimes with preceding punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-968)\nA/1/1030\nStarts of HTML tags: \"<\" or \"</\", sometimes with preceding punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1030)\nA/1/1645\nHTML tag beginnings \"<\", sometimes with preceding punctuation, predicting \"br\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1645)\nA/1/3556\n\"</\", sometimes with preceding punctuation, beginning HTML closing tags\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3556)\nA/1/4005\nCode: \"><\" in XML?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4005)\nA/1/4081\nCode: HTML Angle Brackets before UPPERCASE\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4081)\nA/1/279\n\"<\" in chat logs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-279)\nCluster #174\nA/0/284\nThis feature fires on HTML, CSS, and other code-heavy syntax like indentation, brackets, semicolons, and hash symbols.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-284)\nA/1/953\nCode: CSS preprocessor?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-953)\nA/1/1449\nCode: Newline +/- spaces in CSS\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1449)\nA/1/2253\nCode: image file path manipulation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2253)\nA/1/2635\n\"_\" for snake_case within androidmanifest.xml files\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2635)\nA/1/3003\nCode: CSS padding\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3003)\nA/1/3218\nCode: Punctuation in HTML properties\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3218)\nCluster #215\nA/0/265\nThis feature fires on the opening angle bracket '<' symbol, especially when it is at the beginning of a line or followed by a tag name, suggesting it is identifying HTML/XML tags.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-265)\nA/1/607\nCode: \"</\" in XML (with <string>, <key> tags)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-607)\nA/1/1720\nCode: Token ends with HTML open tag character\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1720)\nA/1/2540\nOpening quotes in technical writing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2540)\nA/1/3056\n\"<\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3056)\nA/1/3955\n\"<\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3955)\nA/1/2423\nCode: Punctuation near packages and versions, esp. in Node\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2423)\nA/1/2416\nCode: opening delimiters and/or quotes of any kind\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2416)\nA/1/1764\n\"=\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1764)\nCluster #231\nA/0/133\nThe feature seems to fire on code in various programming languages, particularly around string literals and function or variable names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-133)\nA/0/177\nThis feature seems to fire most strongly for open/close parentheses.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-177)\nA/0/409\nThis feature seems to fire on punctuation such as quotes, parentheses, brackets, commas, colons, semicolons, arrows, and hyphens. It also fires a bit on dollar signs and formatting such as newlines and tabs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-409)\nA/1/39\nCode: opening quotes in data?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-39)\nA/1/918\nCode: punctuation near price calculations/processing?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-918)\nA/1/1344\nCode: Opening parenthesis in C-like for/if/while statements\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1344)\nA/1/1616\nCode: \"_\" used in snake_case to join multiple words together, usually in code contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1616)\nA/1/2179\nCode: usually some opening delimiter + some quote, often handling basic website user info (PHP/Perl?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2179)\nA/1/2224\n\"[\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2224)\nA/1/2262\nCode: \" $\" as variable sigil\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2262)\nA/1/2274\n\"=\\\"\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2274)\nA/1/3093\nCode: \"(\" calling a function/constructor\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3093)\nA/1/3864\n\" @\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3864)\nA/1/3210\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3210)\nA/1/3019\nCode: Punctuation near file handling\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3019)\nA/0/168\nThis feature seems to fire on code snippets and paths, looking for things like class names, function names, file extensions, #includes, and library imports.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-168)\nA/1/2087\nCode: punctuation in authentication-handling (PHP?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2087)\nA/1/1335\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1335)\nA/1/925\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-925)\nA/1/339\n\"_\" within citation-ish things, often followed by a year-of-publication\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-339)\nA/1/2310\nCode: punctuation near multiprocessing and scheduling\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2310)\nA/1/2708\n\"::\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2708)\nA/1/1069\nCode: time manipulation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1069)\nA/1/172\nCode: Network programming?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-172)\nCluster #232\nA/0/10\nThis feature attends to programming code. The feature fires more for syntax common in programming languages such as curly braces, parentheses, semi-colons, arrows, commas, hashtag/number comments, indentation, variable declarations, function calls, if/else statements, html tags, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-10)\nA/0/462\nThis feature detects hyphens between words that indicate compound terms.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-462)\nA/1/1546\nCode: :: token in C++ namespaces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1546)\nA/1/1846\nCode: C for low-level (circuit boards etc.) programming?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1846)\nA/1/2231\n\"->\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2231)\nA/1/2748\nCode: Underscores in snake_case identifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2748)\nA/1/3747\nCode: Open parentheses, sometimes with more punctuation, in C-like languages\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3747)\nA/1/1669\n\"_\" for snake_case within mathematical variable names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1669)\nA/1/1007\nCode: security key references\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1007)\nA/0/199\nThis feature appears to fire on tokens related to computer networking, such as IP addresses, TCP, TLS, SSL, HTTPS, ports, sockets, servers, certs, URLs, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-199)\nA/0/216\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-216)\nA/1/1450\nCode: mostly punctuation in maybe function metaprogramming?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1450)\nA/1/28\nCode: making/manipulating HTTP requests and responses\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-28)\nA/1/1709\nCode: \" __\" in special identifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1709)\nA/1/3270\nWords related to logging in/authentication\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3270)\nCluster #141\nA/0/358\nThis feature seems to attend to code snippets and programming-related text, looking for things like variable names, function names, keywords, and other code-specific syntax.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-358)\nA/1/126\nCode: Underscores in UPPER_SNAKE_CASE identifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-126)\nA/1/209\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-209)\nA/1/477\nCode: \"_\" in UPPER_SNAKE_CASE identifiers (registers, low-level C?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-477)\nA/1/896\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-896)\nA/1/1201\nCode: CMake\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1201)\nA/1/2328\n\" \\\"\" for opening (and sometimes \".\" for closing) an all-caps quoted sentence, that guides further all-caps\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2328)\nA/1/3347\nCode: _ in C++ macro, especially after BOOST\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3347)\nA/1/2444\nCode: JavaScript or web frameworks (React etc.?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2444)\nA/1/575\nCode: CSS\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-575)\nA/1/2009\nCode: Command lines / shell scripts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2009)\nA/1/2473\nCode: C++? Gtk/UI programming?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2473)\nA/1/3100\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3100)\nA/1/644\nCode: C++ include statements\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-644)\nA/1/371\nEnglish descriptions of character classes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-371)\nCluster #118\nA/0/207\nThis feature seems to fire on code-like tokens such as variable names, function names, keywords, preprocessor directives, nested blocks, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-207)\nA/1/1102\nCode: ALL_CAPS with underscores\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1102)\nA/1/1325\nALLCAPS Syscall names, predicting runs of spaces?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1325)\nA/1/1455\nCode: Uppercase letters in SNAKE_CASE identifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1455)\nA/1/2133\nCode: Uppercase letters in SNAKE_CASE identifiers (in C preprocessor?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2133)\nA/1/3485\nCapitalized letters and punctuation near/predicting \"P\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3485)\nA/1/3589\nThis feature seems to attend to proper names or acronyms starting with the letter M.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3589)\nA/1/3651\nTitleCase word prefixes predicting -ex, esp. in code contexts, esp. \"Mut\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3651)\nCluster #194\nA/0/257\nThis feature fires when it sees web addresses and URLs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-257)\nA/0/262\nThis feature seems to fire on stylistic features such as backticks (`) and different kinds of brackets/parentheses including angle brackets (<>), square brackets ([]) and curly braces ({}). It also fires on short hyphens/dashes (-) and tildes (~).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-262)\nA/1/44\nUsernames in a chat transcript of some sort?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-44)\nA/1/105\nWords in URLs (excluding the \"/\" delimiters), focusing more on characters than numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-105)\nA/1/360\n\"com\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-360)\nA/1/540\nCode: Various kinds of markup code (emails, html, cssy-stuff, misc data formats, etc)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-540)\nA/1/1458\nTLDs in URLS (com in github.com, org in wikipedia.org)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1458)\nA/1/2089\nParts of domain names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2089)\nA/1/2620\nPath components after and predicting \"/\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2620)\nCluster #198\nA/1/289\n\"https\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-289)\nA/1/2370\nCode: Linux command line\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2370)\nA/1/3011\n\" lib\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3011)\nA/1/3129\n\"http\" in urls\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3129)\nCluster #148\nA/0/384\nThis feature seems to fire strongly on uppercase, camel case and punctuation such as hyphens. These all appear to be common in code, variable names, function names, etc. So this feature likely attends to stylistic features common in programming languages and code.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-384)\nA/1/769\nOCR Errors\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-769)\nA/1/2774\nText or word fragments, maybe medical\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2774)\nA/1/3010\n\" extends\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3010)\nA/1/3797\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3797)\nA/1/3085\nNouns in computer programming, which maybe hold code/state\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3085)\nA/1/2121\nObject-oriented programming (English prose)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2121)\nA/1/938\nCapitalized New Word Prediction\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-938)\nA/0/150\nThis feature fires on variable, function and class names in a variety of programming languages.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-150)\nA/1/1263\nThe feature appears to fire on tokens that typically contain the letter v. The feature fires most strongly on occurrences of the letter v itself, but also activates some for words containing v, variable and function names including v, programming language identifiers starting with v, urls containing /v/, and hex color codes containing v.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1263)\nA/1/3367\nCode: identifier fragments esp. non-English, or sometimes typos\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3367)\nCluster #202\nA/0/256\nThe feature fires when it sees code comments, denoted by /* ... */ or // in programming languages such as C, C++, Java, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-256)\nA/1/34\nCode: Comments: Nouns in systems/database engineering?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-34)\nA/1/577\nCode: documentation in markdown?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-577)\nA/1/1600\nCode: Capitalized words at beginnings of code comments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1600)\nA/1/1672\nCode-related English prose, e.g. API documentation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1672)\nA/1/3382\nDocumentation around memory/CPU concepts (stack, pointer, structure, clock, size, etc)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3382)\nCluster #201\nA/0/362\nThis feature attends to code snippets, specifically Python code.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-362)\nA/1/751\nGeneric part of variable name\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-751)\nA/1/3848\nCode: Names being imported in Python\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3848)\nA/1/3884\nCode: Python imports\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3884)\nA/1/4089\nComponents of period-separated attribute/package paths\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4089)\nA/1/2147\nCode: C style comments about low-level hardware\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2147)\nA/1/3444\nTitleCase identifier components, esp. final ones, esp. with some abstract code meaning like \"Class\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3444)\nA/1/692\nCode: predicting byte/buffer?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-692)\nA/1/724\n\"include\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-724)\nA/1/567\n\"import\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-567)\nCluster #65\nA/0/386\nThis feature fires for numbers enclosed with symbols such as parentheses or brackets.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-386)\nA/1/1642\nCaesar-shift encoded words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1642)\nA/1/2950\nASCII control characters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2950)\nA/1/2389\n\"x\" in \"0x\" hex literals with many leading 0s\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2389)\nCluster #107\nA/0/119\nThis feature attends to hexadecimal strings. It appears to fire most strongly on strings of length 6-8.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-119)\nA/1/97\nLong strings of digits\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-97)\nA/1/198\nLowercase hexadecimal (digit token)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-198)\nA/1/1444\nHexadecimal, esp. lowercase letters, predicting digits\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1444)\nA/1/2025\nDigits esp. after the decimal point in decimals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2025)\nA/1/2048\nList of Numbers (space-separated)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2048)\nA/1/3629\n\"x\" in hexadecimal literals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3629)\nA/1/3817\nHexadecimal (uppercase)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3817)\nCluster #64\nA/0/45\nThis feature fires when it sees base64 encoded text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-45)\nA/1/1493\nSynthetic math: repetitive letter strings being picked \"without replacement from\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1493)\nA/1/1544\nbase64 of ASCII\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1544)\nA/1/2357\nbase64 (esp. uppercase letters)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357)\nA/1/2364\nbase64 digits\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2364)\nA/1/1961\nLetters in short fragments of uppercase hexadecimal or some alphanumeric IDs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1961)\nCluster #106\nA/0/283\nThe feature attends to numbers containing hyphens commonly found in phone numbers, dates, times, and item numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-283)\nA/1/327\nHour parts of timestamps, esp. with leading space and 0\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-327)\nA/1/1881\nNumbers in various contexts e.g. error IDs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1881)\nA/1/2215\nGrant and project IDs, mostly digits with few capital letters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2215)\nA/1/3807\n2-digit time components separated by \":\"s\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3807)\nA/1/4027\nNumbers in hyphen-separated sequences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4027)\nA/1/3941\nParts of phone numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3941)\nA/1/1036\nDigits in URLs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1036)\nA/1/2859\nDot-separated version numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2859)\nA/1/1573\nFile path components esp. related to installable software, packages, fonts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1573)\nA/1/605\nSpace+number tokens in various technical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-605)\nA/1/3920\nCode: Punctuation esp. at end of first line of class declarations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3920)\nA/1/2945\nCode: Objective C method declarations, with \":(\" indicating a type will come next as part of a declaration\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2945)\nA/1/3020\nPython-adjacent text/punctuation, esp. \"#,\" preceding \" python-format\" in .po/.pot files\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3020)\nA/1/2998\nUsernames in irc-style <username> format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2998)\nA/0/153\nThis feature attends to HTML tags.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-153)\nCluster #150\nA/0/0\nThis feature attends to HTML and XML tags.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-0)\nA/1/173\nCode: HTML attribute names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-173)\nA/1/259\nCode: Numbers/hexadecimals inside quoted strings for UX-related config/markup (XML?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-259)\nA/1/978\nCode: HTML class name, particularly abstract layout-related like \"container\" or \"child\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-978)\nA/1/985\nAndroid XML layout file properties\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-985)\nA/1/1082\nCode: HTML/XML tag names in closing tags\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1082)\nA/1/1713\nCode: Java/C++ types inside a template/generic instantiation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1713)\nA/1/3457\nCode: HTML/XML opening tag names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3457)\nA/1/4031\n\" class\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4031)\nA/1/4086\nHTML tag names (generally after \"<\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4086)\nA/1/1879\nText in HTML or XML tags (predicting close tags)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1879)\nA/1/3034\nCode: Parts of strings in compact JSON\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3034)\nA/1/4069\nHTML/CSS tutorials (prose)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4069)\nCluster #151\nA/0/266\nThis feature attends to HTML programming code including HTML tags and attributes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-266)\nA/1/244\nMid-word quoted text, predicting closing quotes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-244)\nA/1/1124\nDouble-quotes string as first argument to a function or method call\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1124)\nA/1/2175\nCode: Single-quoted strings (often in an array, set, or function call)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2175)\nA/1/3226\nWords in string literals in JSON or similar in package metadata (\"integrity\", \"metadata\", \"outputs\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3226)\nA/1/1603\nCode: jQuery methods\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1603)\nA/1/2095\nShort text in double quotes, esp. in technical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2095)\nA/1/815\n\" click\" / instructions for interacting with UIs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-815)\nCluster #192\nA/0/209\nThis feature attends to the syntactic and structural features of SQL code. It fires on common keywords, symbols, and formatting in SQL such as SELECT, FROM, DISTINCT, *, \", (, ), \\n, AS, etc. as well as code structure like indentation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-209)\nA/1/1618\nLetter strings that predict \"ID\"/\"id\" (no space)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1618)\nA/1/2060\nVerbs and adjectives in discussions of tabular databases / spreadsheets\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2060)\nA/1/2164\nCode: Testing-related\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2164)\nA/1/3518\nCode: Visual Basic for Excel\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3518)\nA/1/3550\nCode: SQL-related C/C++/Java code, constants, and comments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3550)\nA/1/4010\nCode: SQL\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4010)\nA/1/3825\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3825)\nA/1/2038\nCode: Error conditions esp. \"Failed\" in error message string literals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2038)\nA/1/1242\n\" var\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1242)\nA/1/3012\nCode: .Net Manifest XML?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3012)\nA/1/1269\nCode: related to IDs, users, accounts, and similar\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1269)\nA/1/2779\nCode: Java CamelCase identifier constituents? (\"Manager\" \"Application\" \"Bean\" etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2779)\nA/1/3072\nCode: Field names in JavaScript\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3072)\nCluster #240\nA/0/39\nThe feature attends to code-like structures such as identifiers, keywords and symbols and longer code-fragments in various programming languages.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-39)\nA/0/211\nThis feature seems to attend to UI layout code for mobile apps, firing strongly on class names, methods, and UI element names commonly used when defining app layouts and styles.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-211)\nA/1/236\nCode: \"get\" in getter method names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-236)\nA/1/668\nCode: CamelCase identifier components esp after \"get\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-668)\nA/1/1026\nCode: File Manipulation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1026)\nA/1/1140\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1140)\nA/1/1318\nCode: the start of a line of indented code that is operating on a object/class (vs control flow)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1318)\nA/1/1749\nCode: Java class/interface names in a class definition\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1749)\nA/1/1766\n\"nbsp\" and some other HTML ampersand escape names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1766)\nA/1/1844\nCode: Date/time-related in Java/Objective-C languages\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1844)\nA/1/1876\nCode: Android UX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1876)\nA/1/2015\nCode: CamelCase versions of UI/UX concepts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2015)\nA/1/2107\nCode: Objective C function/method calls on iOS\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2107)\nA/1/2532\nCode: relating to processes and tasks and async execution\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2532)\nA/1/2607\nCode: words in CamelCase, particularly prepositions: \"To\", \"For\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2607)\nA/1/2886\nCode: C# components of CamelCase identifiers (metaprogramming?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2886)\nA/1/4012\nSpanish and Portuguese: Pseudocode or coding related words?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4012)\nA/1/2892\nCode: abstract-ish identifiers in object names, method calls, variable declarations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2892)\nA/1/3228\nCode: $-sigil-prefixed variable names (PHP/Perl?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3228)\nA/1/232\nCode: some serialization or dump of networking info?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-232)\nA/1/1606\n\"Exception\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1606)\nA/1/1227\nCode typically near string concatenation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1227)\nA/1/2417\nThis feature seems to activate for code snippets and programming-related text. Some indications of this: - Frequent use of programming-related symbols and keywords like :=, void, int, def, public, etc. - Indentation and structure common in code, like braces { }, indentation, newlines, etc. - Frequent variable and function names like input, printf, copy, assign, etc. - Presence of comments like /* ... */ and // So in summary, this feature strongly attends to programming code and related text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2417)\nCluster #243\nA/0/388\nThe feature appears to respond strongest to code-like text and variable names, particularly shorter words like \"var\", \"int\", \"mem\", etc. It also seems to fire on function names, boolean values, and math operators.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-388)\nA/1/20\nCode: println and other method calls, esp in Java\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-20)\nA/1/3165\nCode: PHP loop iteratees?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3165)\nA/1/3684\nCode: arguments in comma-separated argument lists\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3684)\nA/1/3644\nCode: \" String\" in C++\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3644)\nA/1/2342\nCode: C++ \" private\" and other keywords\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2342)\nA/1/1149\nCode: Variable names after \"var\", likely JavaScript\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1149)\nCluster #173\nA/0/511\nThis feature attends to code and programming language syntax such as variable declarations and function calls.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-511)\nA/1/186\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-186)\nA/1/454\nCode: punctuation near geometry / graphics calculations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-454)\nA/1/1762\nCode: identifier components used in geometry / graphics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1762)\nA/1/2306\n\" E\" or other identifier prefixes or punctuation predicting \"GL\" in OpenGL contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2306)\nA/1/4080\nThis feature attends to text dealing with lung diseases, specifically chronic obstructive pulmonary disease (COPD).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4080)\nA/1/1995\nCode: Components of CamelCase identifiers (in C/C++?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1995)\nA/1/1476\nCode: Python (Machine Learning) (esp. punctuation)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1476)\nA/1/2193\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2193)\nA/0/349\nThis feature attends to code blocks, particularly Python code. It fires on syntax like variable assignments, function definitions, class definitions, imports, strings, comments, and other Python-specific syntax.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-349)\nA/1/2574\nCode: Tree traversal (mix of Java, Go, Javascript, maybe others)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2574)\nCluster #225\nA/1/430\n\" char\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-430)\nA/1/837\n\" void\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-837)\nA/1/1845\n\" unsigned\"/\"unsigned\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1845)\nA/1/3286\n\" int\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3286)\nA/1/2101\nCode: identifier fragments, often after \"_\", related to data or dimension/shape \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2101)\nA/0/460\nThis feature attends to programming language syntax, particularly C/C++.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-460)\nA/1/386\nCode: C++ template type variables, esp. \" T\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-386)\nA/1/573\nCode: \" std\" and other C++ namespaces (anticipating \"::\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-573)\nCluster #207\nA/0/140\nThe feature appears to fire on identifier names in programming languages. It attends to variable, function, and class names in code.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-140)\nA/0/411\nThis feature fires on source code tokens related to declaring variables, functions, structs, etc. such as int, void, struct as well as punctuation like ; , ( ) { } and identifiers that look like variable names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-411)\nA/1/29\nCode: Beginnings of components of snake_case identifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-29)\nA/1/218\nCode: short variable names?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-218)\nA/1/447\n\" i\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-447)\nA/1/1051\nCode: space+letter tokens, often single-letter variable names)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1051)\nA/1/2993\nCode: C in hardware drivers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2993)\nA/1/144\nCode: C++ type-level programming\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-144)\nA/1/94\nCode: Python Definition\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-94)\nA/1/3992\nCode: \" null\" or sometimes \" false\" esp. after \":\" or \"return\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3992)\nA/1/1760\nFields accessed with member-of-pointer (arrow operator \"->\") in C++\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1760)\nA/1/2468\n\" len\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2468)\nA/1/899\n\" x\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-899)\nA/1/3048\nCode: argument name after \"(\" or \"(&\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3048)\nCluster #92\nA/0/305\nThis feature seems to fire on numbers with decimal points, especially when they are close to 0.5.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-305)\nA/1/705\nCode: Single digits, usually not numeric literals (e.g. in identifiers or versions)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-705)\nA/1/718\nSingle-digit tokens esp. after \"=\", often some software flag or argument\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-718)\nA/1/3989\n\" 0\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3989)\nCluster #242\nA/0/196\nThe feature seems to fire on string formatting in code, particularly left and right parentheses, string formatting characters like %s, and tokens associated with variable names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-196)\nA/1/409\nCode: Identifiers after \"(\", likely arguments, in C#?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-409)\nA/1/1729\nCode: variable in if () condition\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1729)\nA/1/3033\nCode: words referring to measurements or quantities: \"length\"/\"width\"/\"height\"/\"count\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3033)\nA/1/3128\nCode: End of parenthetical\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3128)\nA/1/1442\nCode: Small numeric constants in parentheses (predicting close parentheses)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1442)\nA/1/1379\n\" As\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1379)\nCluster #265\nA/0/282\nThis feature seems to fire when it detects the use of comparative language such as \"than\", \"except when\", and \"other than\". It detects phrases that compare two things, either explicitly using \"than\" or \"more/less than\", or implicitly through use of \"except\" or \"other than\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-282)\nA/1/16\n\" than\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-16)\nA/1/3849\n\" than\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3849)\nA/0/351\nThis feature attends to the word \"as\" used as a conjunction.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-351)\nCluster #264\nA/0/104\nThis feature fires when it sees the word \"as\" being used as a preposition or conjunction.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-104)\nA/1/555\n\" as\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-555)\nA/1/631\n\" as\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-631)\nA/1/1978\n\" as\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1978)\nA/1/3252\n\" as\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3252)\nA/1/3666\n\" as\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3666)\nA/1/963\nAdverb in \"as ... as\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-963)\nA/1/1083\n\"As\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1083)\nA/1/1390\n\" since\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1390)\nA/1/3293\n\" until\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3293)\nA/1/1686\n\" except\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1686)\nA/1/3674\nThis feature fires when transition words like \"furthermore\", \"moreover\", \"thus\", etc are used.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3674)\nA/1/3035\n\" while\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3035)\nCluster #253\nA/1/1111\nExpress uncertainty with maybe/probably/perhaps\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1111)\nA/1/1784\n\" then\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1784)\nA/1/2054\n\" Well\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2054)\nCluster #274\nA/0/502\nThis feature seems to fire on words like \"moreover\", \"because\", and \"however\", which are used to signal logical connections in arguments. So in summary, this feature attends to argument structure and transition words.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-502)\nA/1/622\n\" but\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-622)\nA/1/1560\nSentence-starting word like \"Moreover\" and \"Therefore\" (predicting \",\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1560)\nA/1/3525\n\" however\" mid-sentence (after commas or semicolons)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3525)\nCluster #275\nA/0/29\nThis feature fires when the word \"if\" is present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-29)\nA/1/592\n\" if\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-592)\nA/1/2278\n\"if\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2278)\nA/1/2406\n\"If\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2406)\nA/1/3059\n\"though\"/\"although\"/\"while\"/forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3059)\nA/1/3852\n\" Whether\"/\" whether\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3852)\nA/1/3918\n\" when\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3918)\nCluster #263\nA/0/116\nThis feature fires in response to transition words like \"however\", \"but\", \"although\", \"though\", \"even though\", \"yet\", and \"until\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-116)\nA/0/191\nThe feature fires for the word \"And\" as a discourse or stylistic marker to conjoin phrases or clauses.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-191)\nA/1/355\n\" Or\"/\"Or\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-355)\nA/1/1021\n\" but\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1021)\nA/1/2088\n\" Nor\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2088)\nA/1/2972\nSentence-starting conjunctions \" And\"/\" But\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2972)\nA/1/3304\n\" So\"/\"So\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3304)\nA/1/2531\n\" because\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2531)\nA/1/2699\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2699)\nCluster #269\nA/0/141\nThis feature attends to phrases that contain the word \"that\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-141)\nA/1/254\n\" that\" Relative Pronoun, news\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-254)\nA/1/450\n\" that\" Relative Pronoun, Scientific\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-450)\nA/1/604\n\" that\" Relative Pronoun, Narrative\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-604)\nA/1/1177\n\" that\" conjuction scientific\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1177)\nA/1/1250\n\" that\" Indicative/Conjunction\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1250)\nA/1/1551\n\" that\" Demonstrative Pronoun\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1551)\nA/1/1863\n\" which\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1863)\nA/1/2645\n\" that\" legal\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2645)\nA/1/4082\n\" which\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4082)\nA/0/127\nThis feature appears to activate when surrounding text contains relative pronouns like \"which\", \"who\", and \"whom\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-127)\nCluster #258\nA/0/222\nThis feature attends to the interrogative pronoun \"what\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-222)\nA/1/372\n\" how\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-372)\nA/1/752\n\"Why\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-752)\nA/1/2102\n\" what\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2102)\nA/1/2603\n\" What\"/\"What\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2603)\nA/1/3555\n\"How\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3555)\nA/1/1362\n\" That\"/\"That\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1362)\nA/1/1484\n\" Oh\"/\"Oh\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1484)\nA/1/501\n\" who\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-501)\nA/1/263\n\" Which\"/\"Which\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-263)\nA/1/3288\nAffirmative utterances: \"Yeah\", \"Yes\", \"Okay\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3288)\nCluster #177\nA/0/186\nThis feature is firing for interrogative sentences starting with “what”.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-186)\nA/1/130\n\" can\"/\" could\"/\" would\" after \"How\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-130)\nA/1/156\nWord after \"What\" where \"what\" is the start of a phrase\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-156)\nA/1/1580\n\" you\" as a question (often a yes/no question)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1580)\nA/1/2122\nConcepts/words in questions, predicting the end of the question \"?\", optionally with more punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2122)\nA/1/3626\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3626)\nCluster #170\nA/0/122\nThe feature attends to statements or phrases using the words \"there is,\" \"there are,\" or \"here\" to indicate existence, presence, location or emphasis.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-122)\nA/0/394\nThis feature fires when the token is \"it's\" or \"It's\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-394)\nA/1/141\n\" it\"/\" It\" in scientific contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-141)\nA/1/600\n\" it\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-600)\nA/1/618\nAbstract nouns in \"If\"/\"When\" sentences, esp. computing or astrophysics related (e.g. \"algorithm\", \"universe\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-618)\nA/1/794\n\" it\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-794)\nA/1/1610\n\" you\" and other pronouns after \"thought\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1610)\nA/1/1776\n\" it\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1776)\nA/1/1870\n\"It\" in \"It's\", at the start of quoted speech\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1870)\nA/1/2168\n\" There\"/\"There\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2168)\nA/1/2546\n\" there\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2546)\nA/1/3152\n\" It\"/\"It\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3152)\nA/1/3460\n\" Here\"/\"Here\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3460)\nA/1/3366\nEmphasis, specificity, or exclusivity\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3366)\nCluster #189\nA/0/24\nThis feature attends to first person plural pronouns.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-24)\nA/0/73\nThis feature fires when the word \"you\" or variations of it (\"your\", \"you've\", etc.) are present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-73)\nA/0/171\nThe feature fires when it sees second person pronouns such as \"you\", \"your\", and \"yours\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-171)\nA/0/190\nThis feature attends to sentences with the pronoun \"they\", especially in contexts where \"they\" refers to people.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-190)\nA/1/77\n\" authors\" (and capitalization/spacing variants)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-77)\nA/1/486\n\" We\"/\" we\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-486)\nA/1/616\n\" You\"/\"You\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-616)\nA/1/655\n\" they\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-655)\nA/1/1218\n\" You\"/\"You\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1218)\nA/1/1346\n\" you\" in conditional \"if\" sentences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1346)\nA/1/1793\n\" They\"/\" they\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1793)\nA/1/2084\n\" we\", in political punditry / media reviews?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2084)\nA/1/2270\n\" we\", esp. after a comma, in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2270)\nA/1/2493\n\"We\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2493)\nA/1/2568\nWords after \"we\"/\"We\", esp. adverbs indicating time or chronological position (\"firstly\", \"then\", etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2568)\nA/1/2827\n\"they\"/\"you\"/\"didn\" predicting U+2019\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2827)\nA/1/2847\n\" it\"/\" she\"/\" he\" after \"if\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2847)\nA/1/3014\n\" you\" as subject of subclause\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3014)\nA/1/3310\n\" you\" as direct object\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3310)\nA/1/3532\n\" We\" in scientific contexts (royal \"we found\" etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3532)\nCluster #190\nA/0/154\nThis feature attends to the pronoun \"she\" and associated female pronouns (her, herself).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-154)\nA/0/482\nThe feature fires when it sees pronouns such as \"he\", \"she\", \"who\", etc. in the context of the sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-482)\nA/1/650\n\" he\"/\" she\" esp. in legal reports\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-650)\nA/1/745\n\" He\"/\"He\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-745)\nA/1/1342\n\" She\"/\" she\"/\"she\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1342)\nA/1/2069\n\"he\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2069)\nA/1/2549\nThis feature attends to instances of the pronoun \"he\", referring to a male person.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2549)\nCluster #179\nA/0/224\nThe feature detects first person singular pronouns and contractions such as \"I\", \"I'm\", \"I've\", \"I'd\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-224)\nA/0/271\nThis feature attends to the usage of the word \"I\" and related pronouns like \"I'm\", \"I've\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-271)\nA/1/52\n\" I\" present tense\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-52)\nA/1/1725\n\" I\" (or \"I\") at the beginning of a double-quoted statement\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1725)\nA/1/2241\n\" I\", often in technical/troubleshooting contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2241)\nA/1/2624\n\"I\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2624)\nA/1/2640\n\" i\", typically as a pronoun or in \"iirc\" in casual writing or chat\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2640)\nA/1/2715\n\" I\" in a conditional(subjunctive?) sense: \"when I\", \"If I\", \"because I\", etc\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2715)\nA/1/2765\n\" I\" at the beginning of a sentence\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2765)\nA/1/3487\n\" I\", esp. in medical contexts as either the Roman numeral for 1 or start of an abbreviation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3487)\nA/1/3676\n'I' usually after 'but'\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3676)\nA/1/4022\n\" I\" in a proper noun (usually?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4022)\nA/1/4045\n\" I\" past tense\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4045)\nA/1/874\n\" Which\"/\"Which\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-874)\nCluster #223\nA/0/444\nThis feature appears to fire for sentences with mathematical questions or formulas that reference specific terms in sequences, such as the nth term.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-444)\nA/1/201\nSynthetic math: \" replacement\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-201)\nA/1/1350\nThis feature attends to mathematical terms and concepts such as highest common factor, derivatives, prime factors, and integer rounding.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1350)\nA/1/1388\n\" nearest\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1388)\nA/1/2681\n\" prime\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2681)\nA/1/3682\n\" divided\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3682)\nA/1/3922\nSynthetic math: \" first\"/\" second\"/\" third\" before \" derivative\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3922)\nA/1/3167\nNewlines before/after \"True/False\" in math problems\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3167)\nA/1/2834\nSynthetic math: \" is\" in division problems\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2834)\nA/1/653\nNumerical place ('thousands', 'tens', 'hundreds'), predicts digit, from math dataset\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-653)\nA/1/3397\n\" the\" in \"What is the...\" etc. math problems\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3397)\nA/1/3280\n\"-\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3280)\nA/1/735\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-735)\nA/1/4007\nSynthetic math: \":\" in multisets that are being sampled \"without replacement from\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4007)\nCluster #224\nA/0/96\nThis feature seems to detect math-related questions and phrases, like \"What is...\" and \"Find the...\". It looks for mathematical operations and concepts like derivatives, divisors, factors, roots, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-96)\nA/1/485\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-485)\nA/1/777\n\"of\" in a Math context, for denominator/term/derivative \"of\" a number\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-777)\nA/1/1813\n\" What\"/\"What\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1813)\nA/1/2609\n\" is\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2609)\nA/1/2764\n\" without\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2764)\nA/1/3554\n\" prob\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3554)\nA/1/3652\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3652)\nCluster #185\nA/1/930\n\" (\" for multiple choice answers in math; generic \"(\" in math in lower quantiles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-930)\nA/1/1323\n\" ||\" in wikitext tables involving dates\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1323)\nA/1/2214\nSynthetic math: Spaces and closing parentheses before choices in math multiple-choice questions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2214)\nA/1/2729\n\"⁄\" U+2044 in normalized UTF-8 as latin1 mojibake\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2729)\nA/1/3602\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3602)\nA/1/2192\n\" and\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2192)\nCluster #203\nA/0/505\nThis feature fires on arithmetic symbols such as the minus sign and equals sign.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-505)\nA/1/1986\n\" +\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1986)\nA/1/2216\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2216)\nA/1/2308\n\" -\" as minus sign in math\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2308)\nA/1/2349\n\" -\" indicating a negative number in a math problem\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2349)\nA/1/2474\nThis feature seems to fire when mathematical operations and notation are present, such as determining values, calculating probabilities, solving equations, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2474)\nA/1/2500\n\" =\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2500)\nA/1/3991\nVerbs, esp. imperative, in (often synthetic) math\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3991)\nA/1/3693\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3693)\nCluster #191\nA/0/195\nThis feature fires for characters and patterns associated with mathematical expressions and equations such as parentheses, decimal points, division signs, exponents, equals signs, variables, numbers, and equations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-195)\nA/1/782\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-782)\nA/1/3343\n\"?\" after non-words?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3343)\nA/1/3969\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3969)\nA/1/2856\n\" be\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2856)\nCluster #186\nA/0/105\nThis feature attends to numerical formatting such as numbers, mathematical symbols and fractions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-105)\nA/1/1826\n\"](#\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1826)\nA/1/2789\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2789)\nA/1/3188\nBurmese: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3188)\nCluster #219\nA/0/293\nThis feature seems to attend to mathematical equations and formulas containing variables, operations, and numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-293)\nA/1/2277\nSynthetic math: \"Let\" introducing a math equation after a period\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2277)\nA/1/2753\n\" Suppose\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2753)\nA/1/2207\n\").\" as an end of a math formula sentence\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2207)\nA/1/14\nSynthetic math: Open parenthesis in \"Let a(b)...\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-14)\nA/1/3865\nSynthetic math: numbers being exponentiated (after \"**\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3865)\nA/1/1022\n\"sqrt\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1022)\nA/1/1568\n\"**\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1568)\nCluster #220\nA/0/149\nThis feature appears to fire for numbers and math symbols, particularly exponents, in the context of mathematical equations and expressions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-149)\nA/0/233\nThis feature attends to the phrases like 'let x = ...' and other statements of equations, variables, and mathematical expressions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-233)\nA/1/62\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-62)\nA/1/495\nSynthetic math: Single-letter variables or functions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-495)\nA/1/812\nSingle-letter variables in synthetic math esp. polynomials\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-812)\nA/1/847\nSynthetic math: numeric constants before \"*\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-847)\nA/1/855\nThis feature attends to mathematical expressions involving variables.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-855)\nA/1/1471\nSynthetic math: numbers, about to be multiplied, after \"Solve\"/\"Suppose\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1471)\nA/1/2302\nSynthetic math: digits between newlines (problem numbers?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2302)\nA/1/2572\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2572)\nA/1/2662\nSynthetic math: space followed by digit usu. after \"+\" or \"-\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2662)\nA/1/3032\nYear in Month Day Year format dates\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3032)\nA/1/3091\nSynthetic math: numbers after single-letter keys \"x:\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3091)\nA/1/3387\nSynthetic math: number after \"Let x =\" (where x is any single letter)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3387)\nA/1/3400\nLaTeX math-mode-ending \"$\" esp. shortly after \"Let\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3400)\nA/1/3526\nSingle-letter variables in polynomials? (where exponentiation is **)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3526)\nA/1/3568\nThis feature attends to basic arithmetic operations and equations involving addition, subtraction, multiplication and equal signs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3568)\nA/1/3573\n\")\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3573)\nA/1/3584\nSynthetic math: single-letter variables esp. after number and \"*\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3584)\nA/1/3762\nSynthetic math: Single Token \"*\" in problems\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3762)\nCluster #206\nA/0/68\nThis feature appears to fire on mathematical expressions involving fractions, exponents, parentheses, variables, operators, and numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-68)\nA/0/324\nThis feature fires on mathematical expressions and equations involving multiplication, exponentiation and variable names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-324)\nA/1/46\n\"*\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-46)\nA/1/70\nSynthetic math: numbers esp. in sqrt()\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-70)\nA/1/528\n\"(-\" as the beginning of parenthesized negative number in math \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-528)\nA/1/583\nSynthetic math: Variable name inside the parens of a definition of a function in math\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-583)\nA/1/593\nSynthetic math: \"(\" of a sqrt(number) call where number is numbers and not a variable\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-593)\nA/1/781\nSynthetic math?: Numbers and variables in equations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-781)\nA/1/1321\n\")/(-\" or \"/(-\" in mathematical expressions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1321)\nA/1/1689\n\")/\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1689)\nA/1/2743\nNewline-space in some serialization format? (YAML?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2743)\nA/1/2952\nSynthetic math: exponentiation starting with \"**\" with optional \"(\" and \"-\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2952)\nA/1/3095\nSynthetic math: minus sign in mathematical expressions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3095)\nA/1/3276\n\")*(-\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3276)\nA/1/3472\n\"/\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3472)\nA/1/3617\nSynthetic math: numbers in nested fractions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3617)\nA/1/3743\nSynthetic math: Expression after \"Simplify\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3743)\nA/1/3870\n\"/\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3870)\nA/1/332\nxAB hex escapes (without backslashes or other delimiters) near text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-332)\nA/1/842\nLaTeX punctuation after \\sum or \\prod (either subscript or \"\\\" anticipating \"limits\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-842)\nA/1/1564\nConsecutive numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1564)\nCluster #210\nA/0/292\nThis feature fires for mathematical expressions involving integers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-292)\nA/1/315\nSynthetic math: Arithmetic, numbers with small digits, in unusual bases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-315)\nA/1/876\nNumbers, esp. negative, in comma-separated lists\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-876)\nA/1/1599\nSynthetic math: Numbers that we will need to answer whether it is composite/prime or divides other numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1599)\nA/1/2019\nSynthetic math: Numbers or fractions that are to be sorted\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2019)\nA/1/2489\nSynthetic math: problem involving division\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2489)\nA/1/2867\nSynthetic math: Numbers in math problems\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2867)\nA/1/3306\nMathematical sequences (numbers separated by \", \")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3306)\nA/1/3354\nSynthetic math: Numbers in math problems involving rounding\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3354)\nA/1/1441\n\"}{-\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1441)\nA/1/3814\nCode: fileID, maybe numeric values in \"key:value\" more generally\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3814)\nA/1/635\nDouble-pipe delimited data\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-635)\nA/1/1416\n\"**\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1416)\nA/1/810\nNumbers near Japanese and/or \" -->\" arrows\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-810)\nA/1/3305\nBug ids (mostly in URLs but also outside of URLs in context too)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3305)\nA/1/3076\nNumbers written in exponential notation (but not the \"E\", which is predicted)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3076)\nA/1/2687\nMarkdown footnotes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2687)\nCluster #134\nA/0/301\nThe feature seems to attend to headers of different sections in text, it especially activates on section headers containing numbers and horizontal lines/dashes below the headers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-301)\nA/1/693\n\"######\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-693)\nA/1/728\nLots of dashes separating sections in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-728)\nA/1/2180\n\"================================\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2180)\nA/1/2298\nLegal discovery emails, esp. \"@\" in \"ECT@ECT\" (Enron emails)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2298)\nA/1/2761\n\"Introduction\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2761)\nA/1/3341\n\"![\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3341)\nA/1/3514\n\"References\" / other headers surrounded by (double) newlines\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3514)\nCluster #102\nA/0/110\nThis feature appears to attend to numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-110)\nA/1/10\nNumbers after 3-ish newlines (legal contexts?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-10)\nA/1/2504\n1–2 digit tokens, esp. after double newlines, sometimes after other punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2504)\nCluster #135\nA/0/347\nThis feature seems to fire on backslashes, quotes, hyphens, newlines, parentheses and other punctuation. It is particularly attentive to common punctuation around citations, code, equations, lists, tables, headings, and other text formatting.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-347)\nA/1/1513\nEOT at start\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1513)\nA/1/2754\nEOT in middle\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2754)\nA/1/1033\nLong runs of hyphens\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1033)\nA/1/1807\nUS court case numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1807)\nCluster #128\nA/0/87\nThis feature seems to fire when there are code snippets and programming language examples in the text, especially with programming keywords like \"try\", \"catch\", \"method\", \"class\", etc. or symbols like curly braces ({}), tabs/whitespace, backticks (`) and newlines (\\n). There also seems to be a preference for C# and Java examples.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-87)\nA/0/323\nThis feature fires when it detects the presence of a question followed by a new line, indicated by the \\n token. The feature seems to specifically look for questions in a Q&A format, with the question preceded by \"Q:\" and followed by a new line. Some of the strongest activations occur when a question is formatted in this way: Q: \\n [question] The feature also seems to fire on multiple new line tokens (\\n\\n), indicating a double space after the question. While the feature responds most strongly to typical phrasing of \"Q:\" followed by the question, it also activates to some degree for other question-related words like \"Why,\" \"What,\" etc. So in summary, the feature is detecting a specific formatting pattern of questions in a Q&A context, looking for \"Q:\", new lines, and the content of the question.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-323)\nA/1/717\nQ before a : at the start of a sequence\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-717)\nA/1/1726\nTitle-Case metadata-ish nouns like \"Title\", \"Name\", or \"Location\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1726)\nA/1/2740\n\"0\" or some other numbers after <EOT>\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2740)\nA/1/4036\n\"External\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4036)\nA/1/2163\nUS Court of Appeals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2163)\nA/1/377\nLegal docs: Citation numbers in square brackets\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-377)\nA/1/2721\nLegal docs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2721)\nA/1/449\nPeriod in legal docs, especially conflict of interest?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-449)\nCluster #123\nA/1/1927\n\".\", sometimes with preceding punctuation, in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1927)\nA/1/2404\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2404)\nA/1/3231\n\".\" in \"v.\" (legal cases)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3231)\nCluster #96\nA/0/405\nThe feature fires when sentences end with punctuation (e.g. period, question mark, exclamation mark).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-405)\nA/1/1653\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1653)\nA/1/2311\nPeriod in mathematical proofs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2311)\nA/1/2584\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2584)\nA/1/3425\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3425)\nA/1/3036\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3036)\nCluster #122\nA/0/250\nThis feature fires when it sees case law citations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-250)\nA/0/346\nThis feature fires on abbreviations that are followed by periods such as \"S.\" and \"e.g.\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-346)\nA/1/125\n\".\" after \"Dr\" and few other honorifics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-125)\nA/1/741\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-741)\nA/1/1061\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1061)\nA/1/1206\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1206)\nA/1/1432\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1432)\nA/1/1548\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1548)\nA/1/1697\n\".\" in abbreviations for legal citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1697)\nA/1/1783\n\".\" in legal/formal abbreviations?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1783)\nA/1/2373\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2373)\nA/1/2642\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2642)\nA/1/3520\n\".\" after single letters, likely abbreviations (\"i.e.\", \"e.g.\", \"p.\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3520)\nA/1/3950\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3950)\nA/1/4035\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4035)\nA/1/2794\n\".\" after initials of people (in academic contexts / citations?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2794)\nCluster #91\nA/0/376\nThis feature attends to citations of journal publications in the format \"<Last name>, <Initials.>, <Publication name>, *<Volume number>* (<Year>).\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-376)\nA/1/1060\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1060)\nA/1/1343\n\".,\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1343)\nA/1/1438\n\").\" suffixing a year, in a bibliography/citation context\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1438)\nA/1/1608\n\"**]{},\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1608)\nA/1/1640\n\" [**\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1640)\nA/1/2079\n\".\" in abbreviations of chemistry/physics journals?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2079)\nA/1/3796\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3796)\nCluster #69\nA/1/1045\n\" \\u00a7\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1045)\nA/1/1125\nNumbers inside of [**NUM**]{} citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1125)\nA/1/2533\nPage numbers in citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2533)\nA/1/2865\n\" \\u00b6\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2865)\nCluster #78\nA/1/2173\n\" See\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2173)\nA/1/3138\nPatent Numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3138)\nA/1/4006\n\" also\" of \"See also\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4006)\nCluster #81\nA/0/22\nThis feature attends to references to sections and subsections of legal documents, such as statutes, patent numbers, court cases, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-22)\nA/0/169\nThis feature seems to attend to text that shows dates and days of the week, often in news-like contexts such as headlines.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-169)\nA/1/61\nLegal: ends of case numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-61)\nA/1/136\nLegal numbers after \" at\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-136)\nA/1/183\n\" v\" in \"v.\" (legal cases)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-183)\nA/1/789\nNumbers in legal citations, usually after \",\" and predicting a court\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-789)\nA/1/1489\nSection number (after \"section\" or \"§\", esp. after \"Code\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1489)\nA/1/1559\n\"d\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1559)\nA/1/1659\nLegal digit esp. after \"F.\" before \"d\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1659)\nA/1/1761\nSupreme Court case numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1761)\nA/1/1806\n\" Id\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1806)\nA/1/2132\n\" Cir\" Legal Citation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2132)\nA/1/2222\nThis feature fires on parenthetical citations with a year inside such as (2002).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2222)\nA/1/2706\nMonths in dates, predicting days of months\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2706)\nA/1/3749\nSpace and digit after \"X v. Y,\" court case citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3749)\nA/1/3755\n\" supra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3755)\nA/1/4059\nNumerical day of a date (Month day, day Month, Month day'th, etc)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4059)\nA/1/398\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-398)\nA/1/787\n\".\" in \"Fig.\" (predicting a figure number)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-787)\nA/1/1612\n\".\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1612)\nA/1/1512\nSupplementary data in papers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1512)\nCluster #94\nA/0/378\nThis feature fires on the label \"Fig.\" that usually precedes a figure number in academic papers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-378)\nA/1/278\n\" Fig\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-278)\nA/1/1531\n\" Table\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1531)\nA/1/2488\n\" FIG\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2488)\nA/1/3723\n\" Eq\" before an equation reference\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3723)\nCluster #95\nA/0/238\nThis feature attends to citations of figures in academic papers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-238)\nA/1/265\nShort table names (\" S\" in \"Table S1.\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-265)\nA/1/282\n\"\\\"}\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-282)\nA/1/362\n\"Fig\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-362)\nA/1/3070\nFigure numbers (after \"Figure\"/\"Fig\"/...)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3070)\nA/1/370\n\"[\", especially after \"Fig. \"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-370)\nA/1/543\nLetters in figure names that are parenthesized or bracketed, like \"[Figure 4A]\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-543)\nA/1/3675\nMany hyphens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3675)\nA/1/353\nChemical molecular data, showing atoms, bonds, and geometry, near many spaces and \"---\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-353)\nA/1/283\n\"/\" as dimensional units \"x per y\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-283)\nCluster #129\nA/0/465\nThis feature attends to lines and horizontal lines in particular.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-465)\nA/1/2051\nMany hyphens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2051)\nA/1/3868\nWhitespace, often right after EOT\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3868)\nA/1/3880\nLong runs of spaces in medical reports?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3880)\nCluster #26\nA/0/160\nThis feature fires for square or curly brackets [ ] { }. The brackets are mostly used for citations, references, links, math expressions, arrays, lists, and occasionally for emphasis or indicating action.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-160)\nA/1/473\n\"[\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-473)\nA/1/1229\n\" (**\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1229)\nA/1/2949\n\" [\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2949)\nA/1/3291\n\" ([\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3291)\nA/1/702\n_{ and ^{ in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-702)\nCluster #25\nA/0/229\nThis feature attends to parentheses () in the text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-229)\nA/1/457\n\" (\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-457)\nA/1/536\nOpening parenthesis after name or institution, where the parenthesis often contains a date\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-536)\nA/1/929\nOpen parentheses after more than one newline\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-929)\nA/1/1065\n\" (\" in an informal narrative or commentary\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1065)\nA/1/1259\n\" (\" Citation, Reference, or Note\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1259)\nA/1/1491\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1491)\nA/1/2276\n\" (\" medical/epidemiological context with a focus on statistical analysis\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2276)\nA/1/2943\n\" (\" in a math/physics technical paper\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2943)\nA/1/3064\n\"(\" in a legal context referencing a particular clause: ie \"924(c)(1)\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3064)\nA/1/3365\n\" (\" in a legal/judicial context\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3365)\nA/1/3449\n\" (\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3449)\nA/1/3689\nOpen parentheses with no leading space, often in math notation or chemical formulas\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3689)\nA/1/3776\n\" (\" in a medical/biological context with a focus on biological terms/parenthetical-acronyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3776)\nA/1/3844\n\" (\" in a software-development/technical document\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3844)\nCluster #21\nA/0/52\nThis feature attends to acronyms in parentheses.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-52)\nA/0/249\nThe feature attends to parentheticals, which are phrases or clauses set off by parentheses or other punctuation marks.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-249)\nA/1/157\nLetter after opening parenthesis, particularly \"(i.e.)\" or \"(e.g.)\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-157)\nA/1/292\nParenthesized numbers of quantities/units\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-292)\nA/1/768\nShort lowercase word after \"(\", esp. in English prose mid-sentence\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-768)\nA/1/1457\nSingle lower case letter in parentheses\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1457)\nA/1/2204\n\"see\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2204)\nA/1/2248\nYears in parentheses, esp. citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2248)\nA/1/2296\nWords inside parenthetical expressions that could be the end of the parenthetical phrase, predicting punctuation including \")\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2296)\nA/1/2553\nText in parentheses, esp. proper nouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2553)\nA/1/2673\nParenthesized abbreviations (abbreviating the preceding text?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2673)\nA/1/2842\nThis feature activates most strongly when it sees the presence of mathematical functions and equations, signified by characters like parentheses, exponents, division signs, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2842)\nA/1/3088\n\"0\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3088)\nA/1/3096\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3096)\nA/1/3466\nWords inside parentheses, esp. right after open parentheses, in medical/experimental contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3466)\nA/1/4074\nNumbered citations/footnotes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4074)\nA/1/3990\n\" \\u00b1\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3990)\nCluster #138\nA/0/313\nThis feature attends to numbers with decimal points representing measurements such as weight or concentration.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-313)\nA/1/2140\n\" **\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2140)\nA/1/3162\n\"--\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3162)\nCluster #140\nA/0/472\nThis feature fires for numbers formatted with spacing around the decimal points.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-472)\nA/1/59\n\" \\u23ce\\n \"/\" \\u23ce\\n \"/\"\\u23ce\\n \"/\"\\u23ce\\n \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-59)\nA/1/166\nLong runs of whitespace near numbers/statistics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-166)\nA/1/404\n\" \"/\" \"/\" \"/\" \"/\" \"/\" \"/\" \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-404)\nA/1/726\nRuns of whitespace in medical contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-726)\nA/1/1178\n\" \"/\" \"/\" \"/\" \"/\" \"/\" \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1178)\nA/1/1453\nLong spaces next to escaped arithmetic operators, potentially in tables?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1453)\nA/1/3028\n\" =\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3028)\nA/1/3808\n\" ×\" in numeric dimension pairs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3808)\nA/1/2377\n\"---\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2377)\nA/1/1382\n\" \"/\" \"/\" \"/\" \"/\" \" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1382)\nA/1/3504\nCapital letters and/or punctuation predicting 3-digit numbers?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3504)\nCluster #137\nA/0/88\nThis feature seems to fire on numbers, punctuation like parentheses, commas, periods, colons, and mathematical operators like +, -, =.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-88)\nA/1/88\nClosing parenthesis after number predicting run of space\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-88)\nA/1/3567\nPunctuation or tokens preceding \"fn\" referring to footnotes in [](){} citation format\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3567)\nA/1/134\n\" vs\" / \"versus\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-134)\nA/1/133\nHebrew?: mojibake UTF-8 as latin1?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-133)\nA/1/712\n\" *\" when used in sequence with additional \"*\" and \" * tokens, in comments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-712)\nA/1/3040\n\" *\", usually as markdown bold delimiters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3040)\nCluster #139\nA/0/400\nThe feature attends to tables and formatting, specifically numbers, text alignment, and punctuation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-400)\nA/1/1424\nNumbers and punctuation near numbers, often \"%)\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1424)\nA/1/2424\nHuman body or medical supply related, near and predicting long runs of spaces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2424)\nA/1/2781\n\" \\\\\\<\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2781)\nA/1/3885\n\"*\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3885)\nA/0/477\nThis feature appears to attend to text containing various symbols like parentheses, dollar signs, percent signs, asterisks, hashtags, tildes, etc. Often there is math notation or references to figures/tables.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-477)\nA/1/819\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-819)\nCluster #136\nA/0/67\nThe feature fires when it sees numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-67)\nA/1/1510\nDigits, usually 2 or 3, after the decimal point in such numbers separated by many spaces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1510)\nA/1/3467\n\" 0\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3467)\nA/1/3756\nPredicting whitespace after legal citations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3756)\nCluster #89\nA/0/247\nThe feature fires when it sees dates formatted as two-digit to four-digit numbers representing years between 1900-2020.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-247)\nA/1/869\nA year, in talking about history and what happened\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-869)\nA/1/1634\nYears, 19xx or 20xx, often in citations of journals and the like\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1634)\nA/1/642\n\"*\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-642)\nA/1/834\nANOVA and related statistical tests\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-834)\nA/1/66\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-66)\nCluster #131\nA/0/94\nThe feature attends to numbers with decimals in scientific writing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-94)\nA/0/242\nThis feature fires when it sees statistical symbols and formatting, particularly p-values, standard error, and statistical significance indicators.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-242)\nA/1/469\nNumbers in percentages\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-469)\nA/1/709\nDigits after the decimal point after \"=\" in parentheses\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-709)\nA/1/2793\nNumbers after/near \" ±\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2793)\nA/1/3233\n\"%\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3233)\nA/1/3887\n\" C\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3887)\nCluster #90\nA/0/261\nThis feature attends to numbers between 20-50.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-261)\nA/1/224\n\" million\"/\" billion\" big numbers or things measured in such quantities\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-224)\nA/1/235\nVarious European languages: Number text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-235)\nA/1/740\nPage numbers esp. in page ranges\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-740)\nA/1/828\nNumbers after Title-Case nouns suggesting sections of a larger work, like \"Volume\" or \"Issue\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-828)\nA/1/867\nSingle letters, esp. lowercase, immediately after numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-867)\nA/1/954\nInline numbers in English text, esp. >2 digits predicting measurements/units\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-954)\nA/1/1126\nThis feature fires on numbers that contain a 1, 4 or 8 in the tenths or hundredths place.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1126)\nA/1/1209\nDigits in mixed digits/uppercase letter sequences (uppercase base36 maybe)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1209)\nA/1/1862\nSmall single-digit numbers (as Arabic numerals)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1862)\nA/1/1992\nNumbers in English sentences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1992)\nA/1/2026\n~2-digit numbers (ages, page numbers)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2026)\nA/1/2448\nNumbers in Bible verse citations (separated by colons)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2448)\nA/1/2634\nThis feature attends to numbers greater than or equal to 100.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2634)\nA/1/2820\nNumbers after \" in\"/\" on\" etc., predicting months\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2820)\nA/1/3703\nNumeric dollar amounts, after \"$\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3703)\nA/1/3902\nThis feature fires when it sees numbers and letter combinations that appear to be vehicle, product or model numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3902)\nA/1/908\n\"*\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-908)\nA/1/226\n\"^\" in Markdown-style superscripts in chemistry?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-226)\nCluster #110\nA/0/152\nThis feature attends to numerical values and measurements.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-152)\nA/1/1183\nNuclear Magnetic Resonance\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1183)\nA/1/1954\n\" 0\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1954)\nA/1/2100\nSpace + 2 or 3 digit number, inline in English prose (mechanical engineering?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2100)\nA/1/2337\n~2-digit numbers in medical studies\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2337)\nA/1/2394\nNumbers of units after the decimal point\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2394)\nA/1/2812\n\" 1\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2812)\nA/1/2857\nNumbers before acronyms/units in scientific papers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2857)\nA/1/3284\n1 to 3 digit numbers, often scientific measurements/units\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3284)\nA/1/2210\n\"~\" in chemistry Markdown subscripts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2210)\nA/1/533\nThis feature fires when it sees exponents, superscripts, and other mathematical notation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-533)\nCluster #111\nA/0/51\nThe feature fires when it detects scientific notation and chemical formulae.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-51)\nA/1/154\nBadly formatted math operators\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-154)\nA/1/816\n~2-digit numbers in scientific contexts, predicting various units\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-816)\nA/1/1465\nNumbers in Markdown (\"~\") subscripts near single capital letters (chemistry?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1465)\nA/1/2167\nCode: double-asterisk-emphasized numbers and letters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2167)\nA/1/2362\n\" \\u00b0\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2362)\nA/1/1162\n\" (\" Preceding Roman Numeral\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1162)\nA/1/461\nCode: HTML & escape character before lt/gt/nbsp/quot/amp\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-461)\nA/1/2877\nPunctuation in JSON-like data, regexes... predicting single digits\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2877)\nCluster #39\nA/0/466\nThis feature attends to the ampersand symbol (&).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-466)\nA/1/393\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-393)\nA/1/4013\nLaTeX table (delimiters?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4013)\nCluster #38\nA/0/407\nThis feature fires on numeric values, including both integers and decimals.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-407)\nA/1/388\nNumbers in LaTeX tables\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-388)\nA/1/878\nWhitespace esp. near/between LaTeX math mode expressions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-878)\nA/1/1122\nBase-ten numbers, especially decimal, negative, and in comma-separated lists without spaces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1122)\nA/1/2351\nNumbers and some English/abbreviations, separated by and predicting & (LaTeX?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2351)\nA/1/2393\nNumbers in (X,Y) pairs? (coordinate-based plotting?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2393)\nA/1/1670\nPunctuation tokens in math ending with \"(\" and near \"x\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1670)\nA/1/1991\nMath notation, synthetic math operators, or LaTeX near the letter \"w\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1991)\nA/0/287\nThis feature attends to LaTeX code blocks.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-287)\nA/1/488\nNewline-spaces in LaTeX commands\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-488)\nCluster #63\nA/0/146\nThis feature attends to LaTeX math expressions containing fractions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-146)\nA/0/228\nThis feature seems to fire for dollar signs ($) used in mathematical formulas and equations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-228)\nA/0/303\nThis feature attends to mathematical expressions, especially those containing Greek letters, math symbols, and formulas.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-303)\nA/0/340\nThis feature attends to subscripts that use curly braces such as C_{D}.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-340)\nA/0/470\nThis feature attends to subscripts of mathematical variables.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-470)\nA/0/501\nThis feature attends to mathematical equations and formulas that contain exponents.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-501)\nA/1/167\n\"^{\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-167)\nA/1/237\nLaTeX: \"-\" in math mode (subtraction) esp. superscripts or subscripts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-237)\nA/1/330\nLaTeX: \" =\" assigning a variable\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-330)\nA/1/345\n\"{\" in LaTeX after mbox/hbox/text/...\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-345)\nA/1/364\nSynthetic math: operators near \"f\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-364)\nA/1/444\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-444)\nA/1/813\n\"(\" and other opening delimiters after LaTeX \"\\left\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-813)\nA/1/931\nBackslashes predicting \"u\" in Unicode character escapes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-931)\nA/1/1089\nSynthetic math: parentheses and arithmetic near \"c\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1089)\nA/1/1181\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1181)\nA/1/1233\nLaTeX begin inline math\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1233)\nA/1/1256\nPredicting Short Math Variable\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1256)\nA/1/1328\n\"_{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1328)\nA/1/1754\nMath punctuation near/predicting \"n\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1754)\nA/1/1768\nLaTeX subscripts with preceding punctuation, predicting \"i\", \"j\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1768)\nA/1/1923\n\" $\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1923)\nA/1/1955\nLaTeX: \"_\" used for subscript\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1955)\nA/1/1982\n\" $\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1982)\nA/1/2263\nLaTeX: Underscores and similar punctuation, predicting digits\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2263)\nA/1/2316\n\"^\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2316)\nA/1/2405\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2405)\nA/1/3180\n\"}{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3180)\nA/1/3248\nSynthetic math: \"*\" in polynomials etc\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3248)\nA/1/3403\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3403)\nA/1/3529\nOpen parentheses in LaTeX?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3529)\nA/1/3569\nParens, brackets, subscripts, and mathematical operators, predicting a variable \"k\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3569)\nA/1/3632\nMostly capital letter fragments predicting X, Y, Z\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3632)\nA/1/3889\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3889)\nA/1/3952\nOperators and punctuation in math near/predicting \"q\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3952)\nA/1/3976\n\"=\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3976)\nA/1/871\nLetters near \"+\", often but not always scientific notation \"1e+5\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-871)\nA/1/1017\nLaTeX esp. opening braces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1017)\nCluster #58\nA/0/416\nThis feature fires on math equations written in LaTeX.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-416)\nA/1/375\nLaTeX: Curly brace in setting mathbb font\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-375)\nA/1/399\n\"}{\" in LaTeX, esp. array environments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-399)\nA/1/678\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-678)\nA/1/1664\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1664)\nA/1/1105\nPunctuation in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1105)\nA/1/1275\nLaTeX measurements\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1275)\nA/1/1135\nCapital letters or punctuation near/predicting \"G\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1135)\nA/1/1947\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1947)\nA/1/107\nLaTeX math environment names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-107)\nCluster #34\nA/0/312\nThis feature fires on text snippets that contain LaTex syntax. In particular, it appears to be attending to structural elements like section headers, equations, references, and formatting.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-312)\nA/0/316\nThis feature attends to LaTeX math notation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-316)\nA/1/55\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-55)\nA/1/273\n\"]{\" in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-273)\nA/1/679\nLaTeX: \" \\\\\\\"/\"\\\\\\\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-679)\nA/1/943\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-943)\nA/1/1076\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1076)\nA/1/1205\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1205)\nA/1/2381\n\"12\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2381)\nA/1/2778\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2778)\nA/1/3490\n\"{\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3490)\nA/1/3981\n\"was\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3981)\nCluster #47\nA/0/329\nThe feature attends to LaTeX math environments, equations, and symbols.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-329)\nA/0/418\nThis feature fires for LaTeX math mode environments, equations, theorems, tables, figures, algorithms, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-418)\nA/1/113\n\" [*\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-113)\nA/1/1132\nLaTeX equation tags\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1132)\nA/1/1283\nCitations and references after (\\\\[, predicting \\\\])\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1283)\nA/1/1461\n\" (\\\\\\\\[\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1461)\nA/1/2205\nFilenames in maybe Markdown link targets?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2205)\nA/1/3543\n\"{\" esp. in LaTeX labels\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3543)\nA/1/3578\n\" \\\\\\\\[\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3578)\nA/1/3587\n\":\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3587)\nA/1/3660\nMarkup references to tables/algorithms/etc., in brackets before colons\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3660)\nCluster #27\nA/0/21\nThis feature fires when it sees brackets [] around numbers, letters and variable names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-21)\nA/1/396\nCode: Single-digit numbers in square-bracket array indexing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-396)\nA/1/3505\nEnglish in [square brackets]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3505)\nA/1/2094\n\"$-\", sometimes with prior punctuation, joining hyphenated LaTeX-math-and-word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2094)\nA/1/212\n\".\" surrounding sub/sup for subscript/superscript formatting requests\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-212)\nA/1/359\nData Label prefixes: basically of the form: .[]{data-label=\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-359)\nA/1/999\nLaTeX tokens after open square brackets (left ends of intervals, etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-999)\nA/1/3275\nLaTeX formatting\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3275)\nA/1/3863\n\"amsmath\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3863)\nA/1/851\nCode: identifiers (variables, key or field names) in curly braces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-851)\nCluster #292\nA/0/259\nThis feature fires when it sees mathematical symbols and equations such as dollar signs, subscripts, fractions, greek letters, parentheses, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-259)\nA/1/402\n\"}$$\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-402)\nA/1/638\nThis feature fires when equations (written in LaTeX or MathML format) are present in the text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-638)\nA/1/1507\n\"\\\\\\\\]\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1507)\nA/1/2266\nClosing LaTeX math mode\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2266)\nA/1/3912\nThis feature attends to mathematical symbols and formulas like variables, operators, functions, fractions etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3912)\nA/1/1475\n\"$\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1475)\nCluster #291\nA/0/40\nThe feature attends to mathematical notation such as inline formulas, equations, theorems, proofs, symbols, Greek letters, subscripts, and other math syntax.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-40)\nA/1/175\n\" Let\" in/near math or LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-175)\nA/1/879\n\" Then\"/\" then\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-879)\nA/1/3557\nPrepositions near LaTeX math\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3557)\nA/1/3873\n\" that\" LaTeX Scientific\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3873)\nA/1/3646\nNumeric tokens, sometimes after all-caps abbreviations/identifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3646)\nA/1/1474\n'follows' in logic, usually related to code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1474)\nA/1/549\n\" such\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-549)\nA/1/437\nMathematical quantifier, e.g. \" some\"/\" any\"/\" every\" after \"for\", near LaTeX and predicting \" $\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-437)\nA/1/1195\nThis feature attends to mathematical formulas.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1195)\nA/1/1268\n\" Proposition\"/\" Theorem\"/capitalized words for mathematical statements\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1268)\nA/1/936\nPrepositions and relational verbs, esp. at beginnings of sentences in LaTeX/math-related contexts, predicting \" Eq\" or math mode starting\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-936)\nCluster #307\nA/0/281\nThis feature attends to mathematical terminology and notation related to abstract algebra, especially homomorphisms, isomorphisms, and topological spaces.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-281)\nA/1/491\nTopology, abstract algebra\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-491)\nA/1/1661\nAdvanced physics (mostly GR, but some QFT)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1661)\nA/1/2595\nAbstract adjectives in math: \" convex\", \" regular\", \" arbitrary\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2595)\nA/1/3740\nAstronomical bodies and related phrases: stars, galaxies, dwarfs, pulsars, etc\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3740)\nA/1/3847\nAdjectives in math / linear algebra: \" linear\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3847)\nA/1/1932\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1932)\nA/1/2871\nHigh energy physics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2871)\nCluster #308\nA/0/84\nThis feature attends to technical language related to physics and astronomy, looking for words and phrases related to brightness, luminosity, energy, and associated units of measurement.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-84)\nA/1/588\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-588)\nA/1/3711\nGalactic astronomy (prepositions)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3711)\nA/1/1950\n\" between\"/\" through\"/similar prepositions in electromagnetics/quantum physics?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1950)\nA/0/103\nThe feature fires in response to usage of the preposition \"by\" in contexts describing mathematical or scientific processes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-103)\nA/1/1555\nAbstract verbs or endings, as might be done to equations or expressions, in math / theoretical physics? (\" solve\", \" compute\" etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1555)\nA/1/3429\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3429)\nCluster #293\nA/0/239\nThis feature seems to attend to mathematical expressions, formulas, and equations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-239)\nA/1/1626\n\" Note\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1626)\nA/1/2979\nMath nouns, predicting inline LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2979)\nA/1/3090\nThis feature appears to activate when numbers or mathematical symbols appear in the context of equations, matrices, vectors, norms, spaces, parameters, eigenvalues, functions, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3090)\nA/1/3542\n\" theorem\", \" section\", and other parts/units of math papers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3542)\nA/1/3761\nThe feature appears to fire on symbols and terminology associated with mathematical expressions, equations, operators, and concepts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3761)\nA/1/3975\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3975)\nA/1/4054\nAbstract nouns in mathematics after \"the\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4054)\nCluster #295\nA/0/318\nThis feature fires when it sees the word \"the\" in the token immediately preceding a technical term.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-318)\nA/0/341\nThis feature seems to attend to mathematical expressions involving the word \"the\". Many of the highest activations contain math equations, formulas, or terminology with \"the\" present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-341)\nA/1/512\n\" a\" in the context of math, expected to be followed by a math word. E.g. \"a subset\" or \"a Hilbert space\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-512)\nA/1/1078\n\" the\" in physics, esp electromagnetism?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1078)\nA/1/1452\n\" the\" in theory of machine learning / statistics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1452)\nA/1/1638\n\" the\" in math (advanced geometry? algebra?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1638)\nA/1/1652\n\" The\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1652)\nA/1/1780\n\" the\" in astrophysics?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1780)\nA/1/3362\n\" the\" in mathematics/physics/dynamical systems?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3362)\nA/1/2183\nAstronomy, especially solar-system\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2183)\nA/1/976\n\" a\" in physics or other sciences contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-976)\nA/1/3503\nFamous mathematicians, predicting things named after them like \" theorem\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3503)\nA/1/1637\n\" for\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1637)\nA/1/2395\n\" such\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2395)\nA/1/1811\nThe feature fires when it sees scientific terms or concepts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1811)\nCluster #304\nA/0/357\nThis feature attends to words related to Coulomb interactions in a physics context.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-357)\nA/1/128\nPhysics relating to motion and change: speed, rising/falling, developing/unstable/steady, etc\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-128)\nA/1/342\n\" electron\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-342)\nA/1/1558\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1558)\nA/1/2878\nKey actions, operands and logical constructs that are part of statistical analysis / modeling (prose)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2878)\nA/1/3502\n\" with\" in math (esp. geometry?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3502)\nCluster #285\nA/0/315\nThis feature attends to hyphens appearing in scientific writing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-315)\nA/1/216\nAbstract nouns like \"stability\" in numerical systems, astrophysics, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-216)\nA/1/823\nGraph theory (esp. adjectives)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-823)\nA/1/3158\nNouns in math, esp. linear algebra (\"matrix\", \"vector\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3158)\nA/1/3435\nNeuron networks/machine learning (prose)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3435)\nA/1/1164\n\" function\", esp. in biology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1164)\nA/1/541\nPhysics-related noun of a particular flavor, e.g. \"curvature\", \"velocity\", \"lattice\", \"transition\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-541)\nA/1/341\nThis feature attends to scientific terminology related to dynamics, structures, gradients, perturbations, equations, behaviors, and parameters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-341)\nCluster #277\nA/0/16\nThis feature fires on word endings that describe geometric shapes and locations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-16)\nA/0/210\nThis feature fires when it sees technical terms and concepts commonly found in physics and engineering documents, particularly those related to optics and lasers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-210)\nA/1/65\n\" wave\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-65)\nA/1/322\nPhysics (Optics)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-322)\nA/1/2112\n\" horizontal\"/\" vertical\" geometry?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2112)\nA/1/3733\nGraphics processing?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3733)\nA/1/2134\n\" respectively\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2134)\nA/1/118\nStatistics (English text)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-118)\nA/1/272\nAbstract nouns esp related to paths in astronomy\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-272)\nCluster #16\nA/0/63\nThis feature fires when the word \"according\" is followed by the word \"to\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-63)\nA/1/890\n\" according\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-890)\nA/1/1403\n\"According\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1403)\nA/1/2064\n\" according\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2064)\nA/1/1518\nGraph/diagram-related nouns (curves, lines, squares, diagrams)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1518)\nA/1/2744\nFunctions of mechanical systems / devices\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2744)\nA/1/800\n\" invention\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-800)\nCluster #276\nA/0/31\nThe feature attends to text related to semiconductor transistors with respect to their components and fabrication.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-31)\nA/0/129\nThis feature attends to technical language describing the physical structure or form of mechanical objects.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-129)\nA/0/306\nThis feature attends to descriptions of rotating mechanical parts such as shafts, knobs, axles, and gears.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-306)\nA/1/365\nPrepositions and abstract verbs in mechanical engineering? metalworking?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-365)\nA/1/521\nPrinting / recording / photographing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-521)\nA/1/685\n\"relates to\" (and similar phrases) in patent filings. \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-685)\nA/1/915\nHVAC\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-915)\nA/1/1319\nEnglish nouns for mechanical components, predicting numbers (patents?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1319)\nA/1/1742\nSemiconductors / Computer Manufacturing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1742)\nA/1/2127\n\" crystal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2127)\nA/1/2485\nEnglish (Signal Processing)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2485)\nA/1/3074\nMaterial science phrases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3074)\nA/1/3227\nElectrical engineering\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3227)\nA/1/3792\nEngineering devices (circuit, lamp, disk drive, instrument, device)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3792)\nA/1/3901\nAdjectives describing geometries of mechanical connections\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3901)\nA/1/3230\nConcurrency, multiprocessing?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3230)\nCluster #259\nA/0/81\nThis feature fires on text talking about technical concepts related to electronic communication, signal processing, and data transmission.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-81)\nA/1/233\nCPU architecture\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-233)\nA/1/307\nComputer networking concepts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-307)\nA/1/1841\nConnector words in electronic storage/protocols\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1841)\nA/1/2066\nWireless networks\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2066)\nA/1/2096\n\" surface\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2096)\nA/1/1166\nNouns or word fragments in biology, esp. animals or fungi\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1166)\nA/1/3631\n\" needle\" and broadly similar nouns in sewing/crafting\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3631)\nA/1/2907\nAbstract nouns in medicine and physics, often related to filtering\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2907)\nA/1/1301\nGeneric, inspecific terms for mechanical, biomechanical, and electronic components\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1301)\nA/1/1034\nWords in scientific text, esp. atomic chemistry\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1034)\nA/1/3322\n\"aid\" in \"Chinese/Japanese Patent [Application] Laid-Open\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3322)\nCluster #272\nA/0/254\nThe feature appears to fire on words and phrases related to chemical processes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-254)\nA/1/373\nExcerpts from technical documents describing polymer chemistry processes and properties\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-373)\nA/1/2910\n\" with\"/\" containing\" in chemistry\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2910)\nCluster #260\nA/1/241\n\"xe\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-241)\nA/1/385\nPhrases around chemical substances\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-385)\nA/1/1567\nChemical abbreviations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1567)\nA/1/181\nChemistry \" pH\", \" temperature\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-181)\nA/1/3287\n\" results\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3287)\nA/1/3261\nStatistics-related nouns or endings thereof, in \"######\" section titles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3261)\nCluster #312\nA/0/218\nThe feature attends to words and phrases commonly found in sections describing the results of scientific papers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-218)\nA/1/3214\n\" study\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3214)\nA/1/3253\n\" studies\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3253)\nA/1/4076\n\" difference\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4076)\nCluster #311\nA/0/83\nThe feature attends to words referring to different types of cells.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-83)\nA/0/179\nThe feature attends to text describing figures and plots.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-179)\nA/0/321\nThe feature attends to words and phrases related to cell death and injury.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-321)\nA/1/159\nThis feature fires when it sees words related to chemical processes, such as polymerization, oxidation, precipitation, elution, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-159)\nA/1/336\nThe feature fires on tokens related to biological or medical samples.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-336)\nA/1/997\nMostly \" cells\" and similar terms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-997)\nA/1/1070\n\" Expression\"/\" expression\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1070)\nA/1/1144\n\" age\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1144)\nA/1/1434\n\" group\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1434)\nA/1/1981\nThe feature fires when it sees words related to time periods or durations such as 'after', 'during', 'weeks', 'months', etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1981)\nA/1/2510\nThis feature attends to words and phrases related to growth, proliferation, and increase.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2510)\nA/1/2951\nThe feature fires when it sees words related to genes and genetic concepts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2951)\nA/1/3324\nSome nouns esp. in medicine/science, but pretty varied\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3324)\nA/1/3332\nBiomedical techniques\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3332)\nA/1/3470\nBlood test results (as levels or concentrations)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3470)\nA/1/3948\nAbstract plural nouns in broadly technical/scientific contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3948)\nA/1/4071\nThis feature attends to words related to research, diagnoses and treatments for diseases and disorders.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4071)\nA/1/2321\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2321)\nA/1/1412\n\" population\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1412)\nA/1/1840\nThe feature fires when it follows words that end with \"s\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1840)\nCluster #286\nA/0/49\nThis feature seems to fire on words related to measurements, including units of measurement and words describing amounts. For example, it often fires on words like \"temperature\", \"pH\", \"concentration\", \"volume\", \"score\", \"sensitivity\", \"specificity\", \"latency\", \"rate\", \"flow\", \"velocity\", \"weight\", etc. as well as numerical values.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-49)\nA/1/98\n\" count\"/\" phase\" in various contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-98)\nA/1/3772\nMechanical engineering performance\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3772)\nA/1/1086\n\" patients\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1086)\nA/1/3750\n\" disease\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3750)\nA/1/3131\nBiology terms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3131)\nA/1/907\n\" treatment\"/\" therapy\" medical nouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-907)\nA/1/3760\nUnit of time in days/months/years (but not minutes/hours)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3760)\nA/1/1405\nDemographic information stratified by age\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1405)\nA/1/1933\nSynthetic biology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1933)\nCluster #317\nA/0/41\nThis feature attends to wording related to perceptual and cognitive processing in scientific and research writing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-41)\nA/0/145\nThe feature attends to parenthetical information (e.g. numbers in parentheses).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-145)\nA/1/482\n\"%\" or verbs in stats from medical trials?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-482)\nA/1/811\nMedical imaging, esp. abbreviations like \" CT\"/\" MR\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-811)\nCluster #169\nA/0/108\nThe feature seems to be attending to biological and scientific language describing experiments, results, and processes related to cells, including activation, signaling, proliferation, inflammation, expression, phosphorylation, suppression, upregulation, pathways, responses, and treatments.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-108)\nA/1/22\n+ and - in T cell names? (Markdown-superscripted \"^+^\"/\"^-^\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-22)\nA/1/182\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-182)\nA/1/849\n\"**\" markup in biology/medicine?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-849)\nA/1/1237\n\" although\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1237)\nA/1/2519\nVerbs related to biology / cell processes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2519)\nA/1/3442\nBiochemistry\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3442)\nA/1/3533\nPrepositions in scientific/statistical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3533)\nA/1/2358\n\" chronic\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2358)\nA/1/3964\nPregnancy\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3964)\nCluster #318\nA/0/387\nThe feature appears to fire on words related to anatomy, in particular anatomy of the abdominal and chest region. It attends to words referring to surgical procedures in those regions as well as anatomical structures such as arteries, tendons, and organs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-387)\nA/1/970\nSentences relating to cardiovasciular anatomy and related conditions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-970)\nA/1/2269\nAdjectives in neuroscience?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2269)\nA/1/3186\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3186)\nA/1/3201\nPrepositions / abstract verbs in medical/anatomical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3201)\nA/1/3821\nMedicine (especially surgery)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3821)\nA/1/3078\nCancer\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3078)\nCluster #24\nA/0/221\nThis feature attends to numerical measurements including volume and concentration, as well as experimental procedures in biology such as staining and incubation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-221)\nA/1/25\nAntibodies (maybe antibody assays?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-25)\nA/1/1420\nBio/biochem laboratory terminology (samples, buffers, solutions, etc)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1420)\nA/1/1707\nLife sciences company names with \"Technologies\" in them\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1707)\nA/1/1930\n\"%\" in chemistry contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1930)\nA/1/2212\nThis feature attends to scientific notation, separators, and punctuation frequently found in scientific literature such as research papers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2212)\nA/1/3272\nBiological science assay kits\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3272)\nA/1/3540\n\" After\" and similar connectors in biological contexts (cells etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3540)\nA/1/2686\nNeuroscience\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2686)\nA/1/707\nFitness\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-707)\nCluster #310\nA/0/291\nThis feature appears to fire on words related to science and medicine, particularly biochemistry, nutrition, and digestion. It seems to be looking for technical terms and jargon related to health, diseases, treatments, and research, as well as measurements and chemicals associated with these topics.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-291)\nA/1/211\nMedicine delivery in trials?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-211)\nA/1/508\nText about pharmaceuticals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-508)\nA/1/1505\nDiabetes diagnosis\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1505)\nCluster #168\nA/0/431\nThe feature seems to fire on content words, especially nouns, in scientific papers that describe biochemical processes. It is particularly attentive to words related to molecular interactions and binding.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-431)\nA/0/443\nThis feature fires on text related to genome analysis, in particular mentions of genetic sequences, homologous genes, PCR, variants, transcripts, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-443)\nA/1/947\nBiochemical interactions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-947)\nA/1/1552\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1552)\nA/1/3572\nBiology / genetics?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3572)\nA/1/3182\nAdjectives in biology / medicine\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3182)\nA/1/3641\nEnglish abstract nouns about psychology, behavior, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3641)\nA/1/1088\nVirology studies and infectious disease, discussing viruses, medical conditions, and protective measures\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1088)\nCluster #302\nA/0/38\nThis feature attends to biological and biochemical concepts, particularly related to cells.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-38)\nA/1/557\nThis feature attends to references related to biological and biochemical concepts, such as molecules and substances found in the body.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-557)\nA/1/1127\n\"cell\", mostly as token and somewhat as in biology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1127)\nA/1/2440\nThe feature attends to words and concepts related to tumors and cancer.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2440)\nA/1/2525\nAdjectives or word endings like \"infected\", applied to cells, in medicine/biology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2525)\nA/1/2632\n\" gene\"/\" DNA\"/\" RNA\"/\" protein\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2632)\nA/1/2805\n\" Ig\" esp. in antibody names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2805)\nA/1/2837\nBiological and cellular activity/effect\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2837)\nA/1/3237\nCell types/adjectives in biology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3237)\nA/1/3768\n\" kinase\" etc in biology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3768)\nA/1/4092\nThis feature fires on examples related to biological processes, cellular biology concepts, and biochemistry jargon. Some key examples it picks up on include cellular viability, immunoglobulin, metabolism, morphology, and characterization. It also seems attuned to biological methods and techniques like electrophoresis, chromatography, and microscopy.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4092)\nA/1/3611\nNutrition\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3611)\nA/1/287\nThis feature fires when it sees words or word parts related to chemistry, especially organic chemistry, such as enzyme and substrate names, chemical reaction and process names, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-287)\nCluster #296\nA/1/759\n\" brain\", \" liver\", and other organs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-759)\nA/1/2099\n\" immune\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2099)\nA/1/3238\nMedical, organ-related adjectives (pulmonary, renal, cardiac)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3238)\nCluster #289\nA/0/392\nThe feature attends to concepts related to healthcare, education and training. It picks up on words and phrases associated with these topics such as training, nurse, mentor, continuing education, patient, health education, intervention, clinical, physician, treatment, specialist, academic interests, professional organizations, providers etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-392)\nA/1/2005\nHospital / health care administration\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2005)\nA/1/2565\nMental health\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2565)\nA/1/2921\nSurverys and related concepts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2921)\nA/1/3050\nWords after hyphens in hyphenated phrases, esp. in medicine\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3050)\nCluster #298\nA/0/327\nThis feature seems to fire most on Latin words, Latin abbreviations, and Latin scientific names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-327)\nA/1/93\nBiological genera\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-93)\nA/1/861\nEndings of Latin anatomical terms / diseases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-861)\nA/1/1019\nMedical: drug names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1019)\nA/1/1575\nPhrases related to plants and agriculture\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1575)\nA/1/2289\n\"ide\"/\"ine\" at ends of words, often chemistry or medicine\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2289)\nA/1/2486\n\"in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2486)\nA/1/3292\nSkincare\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3292)\nA/1/109\nNames in author lists, particularly ends (predicting \" et\" of \" et al.\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-109)\nCluster #261\nA/0/476\nThis feature attends to chemistry/biochemistry related terms containing substructures like propyl benzyl and sulfonyl groups.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-476)\nA/1/192\n\" hydro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-192)\nA/1/753\nTokens, often ending in \"z\", in long chemical names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-753)\nA/1/2665\n\" methyl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2665)\nA/1/2978\nTokens in long chemical names with parentheses, brackets, and hyphens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2978)\nA/1/3118\nWords in organic chemistry, esp. predicting \"alkyl\" or \"aryl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3118)\nCluster #299\nA/0/188\nThis feature seems to be attending to descriptions of living things, especially animals. It is particularly activated by words having to do with animal behavior, appearance, habitat, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-188)\nA/1/2279\nHorticulture\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2279)\nA/1/2524\nTransitive verbs and prepositions related to biology/marine sciences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2524)\nA/1/2994\nAdjectives/verbs describing animals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2994)\nCluster #290\nA/0/86\nThe feature attends to words or phrases related to relationships, both romantic and familial.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-86)\nA/1/2554\nLegal descriptions of child abuse\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2554)\nA/1/3282\nSex (general) / queer text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3282)\nA/1/2359\n\" oxygen\"/\" carbon\"/chemical elements and similar\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2359)\nA/1/2484\nEmergency and hospital services and phrases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2484)\nA/1/1716\n\" degrees\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1716)\nA/1/471\nEducation, Schools\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-471)\nA/1/3373\nKnitting (+tailoring?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3373)\nA/1/1173\n\" with\" and other connectors near color/texture/shape adjectives\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1173)\nCluster #297\nA/0/486\nThis feature fires on measurements and cooking instructions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-486)\nA/1/424\n\" the\" in recipe instructions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-424)\nA/1/835\n\",\" in recipes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-835)\nA/1/850\nConnector words in recipes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-850)\nA/1/1273\nRecipe Name (Title Case)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1273)\nA/1/1649\nRecipe ingredient lists\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1649)\nA/1/2455\nNumbers, often in fractions, of measurement in recipes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2455)\nA/1/3337\nRecipe text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3337)\nA/1/3706\nFood reviews\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3706)\nA/1/2383\n\" care\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2383)\nA/1/1609\nWords for organisms or living structures, esp. \"bacteria\" or plural\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1609)\nA/1/1179\nWords/tokens, often \" crypt\", often in cellular biology or other sciences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1179)\nA/1/2509\n\" risk\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2509)\nA/1/1085\n\" method\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1085)\nA/1/1157\nThe feature fires when it encounters the word \"These\" at the beginning of a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1157)\nA/1/3783\nThe feature fires when the word \"for\" is followed by a word related to an analytical or scientific purpose such as determination, detection, estimation, measurement, planning, modeling, study, analysis, formulation, measurement, planning, processing, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3783)\nA/1/1295\n\" confidence\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1295)\nCluster #358\nA/0/155\nThis feature seems to fire on words and phrases related to business operations and processes. Some examples include marketing, optimization, management, efficiency, identification, testing, inspection, collaboration, and planning. There is a particular focus on search engine optimization.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-155)\nA/1/1355\n\" benefit\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1355)\nA/1/3386\nCorporate buzzwords\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3386)\nA/1/3606\n\" research\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3606)\nCluster #359\nA/0/442\nThe feature attends to concepts related to techniques, methods, programs and systems.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-442)\nA/1/160\nAbstract nouns in biology and technology related to testing, support, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-160)\nA/1/1174\n\" issues\"/\" problems\" in formal English\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1174)\nA/1/3879\n\" etc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3879)\nCluster #99\nA/0/156\nThis feature attends to the use of the word \"using\" in English text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-156)\nA/1/162\nCode: comments: web programming questions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-162)\nA/1/576\nConjunctions?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-576)\nA/1/877\n\" Some\"/\" some\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-877)\nA/1/1255\n\" example\", often after \"for\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1255)\nA/1/1410\nWords shortly after I / you / aux. verb + \"be\" / contraction + \"have\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1410)\nA/1/1975\nInstallation/build tool instructions that refer to directories\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1975)\nA/1/3585\nVerbs like \"convert\"/\"add\"/\"remove\", often targeting data, esp. present-tense or infinitive\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3585)\nCluster #348\nA/0/352\nThe feature fires when it sees words related to data and information.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-352)\nA/1/2862\n\" material\"/\" content\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2862)\nA/1/3401\nScientific data\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3401)\nCluster #344\nA/0/197\nThis feature attends to descriptions and uses of systems and models.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-197)\nA/1/706\n\" area\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-706)\nA/1/1155\nNouns related to paths, shapes, and movement\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1155)\nA/1/1425\n\" point\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1425)\nA/1/2669\n\" line\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2669)\nA/1/4061\n\" form\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4061)\nA/1/2267\n\" error\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2267)\nCluster #356\nA/0/113\nThis feature appears to attend to mentions of UI elements and programming concepts like functions, objects, variables, tables, and code snippets.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-113)\nA/1/239\nAbstract nouns in prose discussion of code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-239)\nA/1/1680\nProgramming-related nouns esp. pertaining to web requests\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1680)\nA/1/1791\n\" version\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1791)\nA/1/1837\nData format or handling tool after \"from\"/\"in\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1837)\nA/1/2802\n\" file\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2802)\nA/1/3211\nAbstract nouns esp kinds of data or abstractions in code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3211)\nA/1/3746\nAbstract data-holding nouns in programming\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3746)\nA/1/3426\n\" information\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3426)\nCluster #355\nA/1/565\n\" class\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-565)\nA/1/1509\n\" text\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1509)\nA/1/1883\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1883)\nA/1/1945\n\" link\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1945)\nA/1/2674\n\" system\" and similar abstract nouns, esp. in computing contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2674)\nA/1/2104\n\" switch\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2104)\nA/1/2880\n\"photo\", most likely mentions of \"photo\" in caption in the news articles \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2880)\nCluster #343\nA/1/982\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-982)\nA/1/1996\nThis feature fires when there are words related to modification or descriptors such as '-ing' endings. It also activates for words related to creation or development.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1996)\nA/1/3843\nComma-delimited lists of items\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3843)\nA/1/2268\nWords in technical contexts with manual line breaks?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2268)\nA/1/3313\n\" search\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3313)\nA/1/1254\nAbstract provisions, often after \"provide\" or \"has\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1254)\nA/1/2946\n\"attempt\" to do smth \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2946)\nA/1/460\n\" account\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-460)\nA/1/2626\n\" test\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2626)\nCluster #346\nA/0/163\nThis feature attends to words and concepts related to taxation and finance.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-163)\nA/1/206\n\" business\" and related nouns or synonyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-206)\nA/1/2350\n\" money\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2350)\nA/1/2992\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2992)\nA/1/3136\n\" estate\"/\" property\"/...\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3136)\nA/1/3837\n\" parking\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3837)\nA/1/4034\n\" credit\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4034)\nA/1/4040\n\" tax\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4040)\nA/1/2535\nEconomic terms and phrases\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2535)\nA/1/1467\n\" security\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1467)\nA/1/1682\n\" boot\", esp. in computing/OS contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1682)\nA/1/505\nClimate Change\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-505)\nCluster #326\nA/0/204\nThe feature attends to sentences with numbers including percentages and dollar amounts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-204)\nA/1/301\nEconomic discussion of tax and subsidy policy\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-301)\nA/1/3441\nEconomics / Business\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3441)\nA/1/3437\nWord fragments, esp. ends, faintly construction-related (\"ver\" in \"cantilever\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3437)\nA/1/3621\nWord endings in medicine\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3621)\nA/1/4072\nSoftware names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4072)\nCluster #230\nA/0/212\nThis feature attends to long technical terms, units and some domain specific symbols.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-212)\nA/0/217\nThe feature attends to end-to-end workflows that include information about the service provider, specific steps, and goals of the workflow.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-217)\nA/0/436\nThe feature fires when it sees names of technologies and software, such as operating systems, browsers, programming languages, apps and tech companies. For example, it activates on words like Android, Windows, Chrome, Swift, Java, Apple, Microsoft, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-436)\nA/1/48\nSEO tips\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-48)\nA/1/85\nPhysical goods commerce\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-85)\nA/1/129\nFantasy games (e.g.. WoW, MTG)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-129)\nA/1/684\nGambling\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-684)\nA/1/881\nTopics relating to books (both \"cover\" and sketchbooks and guides and records, but also ebooks, paperwhite, and PDFs)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-881)\nA/1/1376\nCPU specs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1376)\nA/1/1656\nSmartphones\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1656)\nA/1/2185\nCybersecurity\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2185)\nA/1/2187\nElectronic devices (cameras, phones, TVs)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2187)\nA/1/2695\n\" web\" (almost always as a single word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2695)\nA/1/2831\n\" email\"/\"mail\" in \"e-mail\" and other mail-related text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2831)\nA/1/2915\nVideo games\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2915)\nA/1/3039\n\" software\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3039)\nA/1/3189\nOperating systems and similar platforms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3189)\nA/1/3193\nContext: offering a sale/discount\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3193)\nA/1/3549\nEnterprise or resume language\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3549)\nA/1/3623\nTechnology hardware review?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3623)\nA/1/3624\n\" computer\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3624)\nA/1/3944\nCryptocurrency\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3944)\nA/1/3701\nSpecifications or ad copy for physical devices / equipment\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3701)\nA/1/519\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-519)\nA/1/2901\n\" figure\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2901)\nA/1/3430\nPhysical objects in construction/architecture/mechanical engineering\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3430)\nA/0/274\nThis feature fires when it sees terms and phrases related to specific parts of the home such as a garage, alley, trailers, ceilings, walls, and floors.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-274)\nA/1/515\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-515)\nA/1/1931\nSentences about automatives and cars\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1931)\nA/1/661\nSome mid-word tokens predicting plural endings like \"ers\", broadly related to hygiene, medicine, firearms, or machinery\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-661)\nA/1/760\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-760)\nCluster #61\nA/0/227\nThis feature appears to attend to various types of medical and biomedical terminology. It activates on words and phrases related to medical conditions, treatments, procedures, diagnoses, anatomy, and other biomedical concepts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-227)\nA/1/946\nNouns, esp. abstract and related to medicine or illness\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-946)\nA/1/1699\nCapitalized words shortly after EOT\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1699)\nA/1/1736\nOnly vaguely software-related verbs, but predicting HTML/JS-related words?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1736)\nA/1/3809\nPlural nouns / endings thereof, vaguely related to tourism\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3809)\nA/1/2071\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2071)\nA/1/151\n\" force\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-151)\nA/1/413\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-413)\nCluster #342\nA/0/72\nThe feature attends to scientific and medical terminology.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-72)\nA/1/410\n\" motor\"/\"Motor\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-410)\nA/1/827\n\" electric\"/\" magnetic\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-827)\nA/1/1040\n\" nuclear\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1040)\nA/1/1397\n\" medical\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1397)\nA/1/1464\nClimate change\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1464)\nA/1/1788\nBiological/genetic adjectives\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1788)\nA/1/2228\n\" health\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2228)\nA/1/2903\n\" visual\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2903)\nA/1/3213\n\" mental\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3213)\nA/1/3777\n\" human\" and some other species/nouns in medicine and social sciences?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3777)\nA/1/3905\n\"supplementary\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3905)\nCluster #233\nA/0/5\nThis feature looks for words that have hyphens in them, as well as some compound words without hyphens like \"offline\" and \"firstname.\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-5)\nA/0/232\nThis feature attends to words ending in \"-able\", \"-er\", \"-ist\" and other similar suffixes that turn a verb or adjective into a noun that performs an action.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-232)\nA/1/27\nWord components, often near word ends, often of professions or \"Xer\" words that mean \"people who X\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-27)\nA/1/335\nThis feature fires when it encounters words that end in -er, -ar, or -or.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-335)\nA/1/376\nThis feature fires when it sees words that end in -off, as well as the word \"link\", \"stop\" or \"screen\". It seems particularly focused on two word phrases where the second word ends in -off, like \"back-off\", \"stand-off\" or \"pick-off\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-376)\nA/1/969\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-969)\nA/1/1501\nThis feature seems to fire for words that contain \"over\" in the middle, especially when they are location names like Hanover-square and Durnover Moor.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1501)\nA/1/1918\nMid-word tokens, esp. in medieval/fantasy contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1918)\nA/1/2317\nWords after hyphens in hyphenated phrases, esp. adjectival phrases?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2317)\nA/1/2769\nText near soft hyphens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2769)\nA/1/3147\nMid-word or ending tokens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3147)\nA/1/3254\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3254)\nA/1/3608\nShort letter sequence token after hyphen in hyphenated phrase\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3608)\nA/1/3680\nThis feature seems to attend to words ending in -er or -en.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3680)\nA/1/3799\n\" social\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3799)\nA/1/40\nEnd-of-word \"ic\", esp. in chemical/biological contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-40)\nA/1/2353\nLarge bodies of water and related words: rivers, coasts, oceans, etc\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2353)\nA/1/2518\n\" olive\" esp in recipes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2518)\nCluster #345\nA/1/677\n\" urban\"/\" rural\", predicting types of areas\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-677)\nA/1/1905\n\" agricultural\"/\" industrial\"/adjectives\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1905)\nA/1/2922\n\" political\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2922)\nCluster #347\nA/0/69\nThe feature seems to fire on words ending with \"ing\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-69)\nA/1/613\nDescriptions of luxurious accommodations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-613)\nA/1/2211\nBuilding, construction, exterior, materials related words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2211)\nA/1/3115\nConnector words in weather/nature descriptions?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3115)\nA/1/3130\nFlowers, trees, descriptions of nature\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3130)\nA/1/3678\nClothing style\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3678)\nA/1/2199\nAdjectives and descriptions of transportation (buses, trains)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2199)\nA/1/4052\nThis feature seems to fire on words that end in \"ins\" or variations of that spelling. Some examples of words it fires on include Martins, Sri Lankans, Cousins, states, rinks, delicacies, injunctions, Alps, draws, jokes, rants, floats, fellows, councils, felis. It seems to specifically detect words ending in \"ins\", \"ans\", \"s\", or similar spellings.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4052)\nA/1/3866\n\" drug\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3866)\nCluster #352\nA/0/379\nThe feature attends to colors, particularly darker colors like black, brown and dark green.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-379)\nA/1/1313\n'dark' token\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1313)\nA/1/2171\n\" wild\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2171)\nA/1/3209\nColors, particularly \" red\" and \" black\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3209)\nA/0/288\nThis feature seems to fire for locations with a geographical focus, specifically rivers, ridges, creeks and valleys.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-288)\nA/1/1338\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1338)\nCluster #375\nA/0/380\nThis feature fires when the text mentions water.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-380)\nA/1/15\n\" fire\" and some similar nouns in police/military/weaponry contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-15)\nA/1/75\n\" gas\", some other forms of power in lower quantiles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-75)\nA/1/271\n\" heart\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-271)\nA/1/442\nBeverages esp. \" wine\", with some food at lower quantiles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-442)\nA/1/462\n\" gold\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-462)\nA/1/615\n\" food\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-615)\nA/1/736\n\" power\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-736)\nA/1/799\n\" blood\" and other medical/bodily fluids\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-799)\nA/1/961\n\" fat\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-961)\nA/1/1698\n\" sleep\", \" pain\", and human conditions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1698)\nA/1/1968\nBody parts, particular on ones with bones or joints or can be injured\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1968)\nA/1/2463\n\" body\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2463)\nA/1/2600\n\" sun\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2600)\nA/1/3000\n\" ball\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3000)\nA/1/3077\n\" air\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3077)\nA/1/3176\n\" mass\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3176)\nA/1/3627\n\" heat\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3627)\nA/1/3741\n\" snow\" and other precipitation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3741)\nA/1/3957\n\" hair\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3957)\nA/1/3987\n\" light\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3987)\nA/1/1225\n\" self\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1225)\nA/1/2141\n\" grass\" and other farm-related concepts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2141)\nA/1/3985\n\" land\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3985)\nA/1/3241\n\" car\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3241)\nCluster #376\nA/0/258\nThis feature attends to words and phrases related to roads, streets, and pathways.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-258)\nA/0/509\nThis feature attends to any mention of a hand, whether in literal hand references like \"he shook my hand\" or figurative ones like \"the stock was in capable hands\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-509)\nA/1/465\n\" mind\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-465)\nA/1/561\n\" door\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-561)\nA/1/647\n\" room\" and similar building/architecture-related nouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-647)\nA/1/1308\n'head' token, most strongly for terms that aren't the body part\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1308)\nA/1/2202\n\"border\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2202)\nA/1/2301\n\" road\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2301)\nA/1/2552\n\" hand\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2552)\nA/1/3763\n\" side\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3763)\nA/1/2040\nTelephones/phones/cell phones\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2040)\nA/1/7\nNouns described by other nouns (locations / \" game\"?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-7)\nA/1/1534\n\" home\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1534)\nA/1/2680\n\" front\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2680)\nA/1/2689\nNouns related to weapons/machinery/transportation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2689)\nA/1/1480\nCardinal directions, e.g. \" north\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1480)\nA/1/1668\n\" step\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1668)\nA/1/3590\nFood-related or celebratory occasions like \" dinner\"/\" tea\"/\" Christmas\"/\" wedding\"/\" birthday\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3590)\nA/1/3114\n\"Welcome\"/\" blog\" and other words in website self-descriptions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3114)\nA/1/2702\n\"cry\" as a verb, as part of compound words, and as part of the idiomatic phrase \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2702)\nA/1/1731\n\" school\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1731)\nA/1/506\n\" city\"/\" town\" and similar location words, esp. discussing historical locations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-506)\nA/1/1113\n'part' token\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1113)\nA/1/493\n\" sound\"/\" voice\"/similar nouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-493)\nA/1/1683\n\" secret\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1683)\nA/1/1012\n\" war\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1012)\nA/1/1804\n\" States\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1804)\nCluster #338\nA/0/183\nThis feature fires on documents involving general concepts around official government and court communication, including words related to reports, hearings, statements, affidavits, letters, and other formal language.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-183)\nA/0/377\nThis feature attends to words associated with government and politics.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-377)\nA/1/1339\n\" family\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1339)\nA/1/1759\n\" team\"/\" crew\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1759)\nA/1/1941\n\" voting\"/\" election\"/\" campaign\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1941)\nA/1/866\nGovernment organizations: Committee, Government, Congress, Legislature, House, Administration\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-866)\nA/0/432\nThe feature appears to fire on proper nouns that are locations such as cities, regions, states, countries, trails, rivers, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-432)\nA/1/451\nCities after \"the\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-451)\nCluster #341\nA/0/97\nThis feature seems to fire for words related to film and cinema, such as movie, drama, TV, television, and video.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-97)\nA/1/962\nMostly single-token \" game\" / associated sports words: \" defense\", \" race\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-962)\nA/1/2145\n\" TV\"/\" television\"/similar forms of media\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2145)\nA/1/2914\n\" book\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2914)\nA/1/3369\n\" news\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3369)\nA/1/665\n\" project\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-665)\nCluster #152\nA/0/433\nThis feature fires when it sees job titles and roles, specifically ending in -er or or. Examples include coordinator, editor, designer, manager, director, writer, analyst, mentor, performer, and adviser.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-433)\nA/1/1005\n\" attorney\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1005)\nA/1/1758\nCode: Strings in smart quotes, esp. in JSON-like formats\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1758)\nA/1/1903\n\" Judge\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1903)\nA/1/1908\nProfessions or endings thereof, esp. writing-related, e.g. \" writer\", \"ologist\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1908)\nA/1/2254\n\" Clerk\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2254)\nA/1/2422\nCapitalized legal parties (\" Defendant\", \"Appellant\", \" Respondent\", \" Plaintiff\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2422)\nA/1/2963\nPolitical Titles, e.g. \" Minister\" or \" President\", predicting politician names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2963)\nA/1/3601\nThe speaker of a phrase, esp. a noun, predicting \" said\" or how they said that phrase\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3601)\nA/1/1771\nTransportation/natural-light-related nouns after \"the\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1771)\nA/1/1604\nBody parts and clothing, most strongly \" chin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1604)\nCluster #280\nA/0/269\nThis feature attends to words associated with male gender such as \"man\", \"boy\", \"boyfriend\", \"father\", \"husband\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-269)\nA/1/1194\nWords for human relatives like \" mother\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1194)\nA/1/1330\n\" user\"/\" customer\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1330)\nA/1/2663\n\"God\"/\" God\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2663)\nA/1/3045\n\" man\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3045)\nA/1/3298\n\" son\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3298)\nA/1/3699\nFantasy roles and characters: \"warrior\", \"angel\", \"hero\", \"vampire\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3699)\nA/1/3980\n\" character\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3980)\nA/1/3049\nBody parts in the context of showing emotion (eyes, cheeks, hands, palms)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3049)\nA/1/2890\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2890)\nA/1/3445\n\" report\"/\" statement\"/synonyms, as released or read by a person or organization in news contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3445)\nA/1/3904\n\" patent\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3904)\nCluster #328\nA/0/148\nThis feature fires when it sees references to time measured in months, years, days, weeks, quarters, or hours.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-148)\nA/1/511\n\" time\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-511)\nA/1/1054\n\" years\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1054)\nA/1/1830\n\" season\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1830)\nA/1/1849\nDays of the week\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1849)\nA/1/1915\nPlural units of time like \"hours\" or \"minutes\" esp. after small numbers or \"few\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1915)\nA/1/2242\nTime of day \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2242)\nA/1/2283\nLarger units of distance\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2283)\nA/1/2882\n\"year\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2882)\nA/1/3469\n\" moment\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3469)\nA/1/3788\n\" year\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3788)\nA/1/3380\nThe feature appears to fire on words related to official meetings, such as meeting, event, launch, conference call.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3380)\nA/1/1700\n\" decision\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1700)\nA/1/3174\n\" comment\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3174)\nA/1/2906\n\" view\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2906)\nA/1/326\n\" case\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-326)\nA/1/2806\n\" bit\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2806)\nA/1/1676\n\" problem\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1676)\nA/1/1270\n\" question\"/\" answer\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1270)\nA/1/887\n\" subject\"/\" object\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-887)\nA/1/690\nAbstract nouns after \"Each\"/\"Every\"/\"One\"/\"Another\"/\"first\"/...\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-690)\nA/1/146\n\" lot\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-146)\nA/1/2941\nThis feature attends to nouns related to information and data such as report, process, data, library, tape, quiz, review, collection, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2941)\nA/1/2594\n\" idea\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2594)\nA/1/1068\nAbstract nouns esp. after \"is/has/have no\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1068)\nCluster #366\nA/0/450\nThis feature fires when it sees sentences that start with \"The\" and are the first sentences of paragraphs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-450)\nA/1/1443\nNouns (inanimate / \"information\"-related?) after \"The\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1443)\nA/1/2670\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2670)\nA/1/3581\nAbstract noun usually after \"This\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3581)\nCluster #360\nA/0/174\nThe feature fires when it sees the term \"here\" and words related to talking about facts or sharing information.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-174)\nA/1/21\n\" way\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-21)\nA/1/2023\n\" place\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2023)\nA/1/2118\n\" reason\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2118)\nA/1/3190\n\" deal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3190)\nA/1/895\n\" fact\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-895)\nCluster #367\nA/0/64\nThis feature fires when the input contains words related to collecting or a collection.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-64)\nA/0/393\nThis feature fires when there are terms related to the importance or significance of something.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-393)\nA/1/73\nAbstract nouns after \"NAME is a/an\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-73)\nA/1/397\n\" name\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-397)\nA/1/1099\nThe feature seems to fire when the word \"the\" is followed by a word related to importance, meaning, nature, science, or power.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1099)\nA/1/1134\nMid-word/name lowercase token after \"A\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1134)\nA/1/1936\n\" kind\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1936)\nA/1/2165\nThis feature appears to fire for abstract nouns which represent a concept, idea or plan. It tends to fire for nouns related to decisions, arguments, secrets, virtues, policies, tactics, recommendations, objectives, signals, efforts, instructions, defenses, assumptions, barriers, distortions and structures.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2165)\nA/1/2757\nVaried nouns after \"a\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2757)\nA/1/3355\nAbstract nouns about choices/plans esp. after adjectives with strong valences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3355)\nA/1/3700\n\" couple\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3700)\nA/1/3393\nThis feature fires when it detects phrases describing measurements like size, length, rate, concentration, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3393)\nA/1/2593\nWord after \"each\" with \"each thing\" semantic meaning\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2593)\nCluster #321\nA/0/420\nThis feature attends to mentions of presence or absence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-420)\nA/1/1098\nAbstract nouns in project management and logistics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1098)\nA/1/1231\nAbstract nouns, esp. quantity-related, after \"a\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1231)\nA/1/1383\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1383)\nA/1/1922\n\" number\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1922)\nA/1/2078\n\" basis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2078)\nA/1/2671\n\" amount\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2671)\nA/1/3510\n\" role\" in \" play a [...] role\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3510)\nA/1/3560\nThe feature fires when it sees terms and phrases related to presence or absence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3560)\nA/1/3729\n\" price\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3729)\nA/1/3881\n\" ability\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3881)\nCluster #217\nA/0/302\nThis feature seems to attend to both descriptions of the face (eyes, mouth) and emotions (tears, smile, grin).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-302)\nA/1/1385\nStanding/walking/climbing and associated adverbs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1385)\nA/1/3026\nThe feature fires when it sees words related to emotional distress, struggling, pain, and hardship.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3026)\nA/1/3475\n\" frowned\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3475)\nA/1/3663\nPhysical / human-body-related verbs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3663)\nA/1/3973\n'shook her head' token, probably gender-neutral\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3973)\nA/1/4042\nThis feature seems to be firing for words related to things you do with your hands such as reading, tying, drawing, and rowing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4042)\nA/1/628\nThis feature seems to attend to things like solutions, choices, behaviors, opportunities, strategies, moves, arguments, and interactions. In general, it seems to fire on concepts related to decision-making, problem-solving, and key actions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-628)\nCluster #350\nA/1/755\nAbstract plural nouns, broadly related to games/music/aesthetics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-755)\nA/1/1723\nThis feature attends to mentions of towns, cities and other geographic areas.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1723)\nA/1/3738\nAbstract nouns for forms of text/images/media\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3738)\nA/1/2177\nPlural noun after a number\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2177)\nA/1/2752\ntheoretical constructs, elemental concepts and functional dynamics that make up larger philosophical ideas/ frameworks \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2752)\nA/1/247\nAbstract nouns indicating relations between usually two things\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-247)\nA/1/453\nPhysical nouns related to nature or less modern construction\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-453)\nCluster #340\nA/0/57\nThis feature fires on words and phrases related to unpleasantness, aggression, suffering, conflict, and mortality.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-57)\nA/1/630\nMostly \" death\" and related terms, but also \" birth\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-630)\nA/1/1246\nThe feature seems to fire most strongly when proper nouns related to military, governments, or political groups/movements are present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1246)\nA/1/1561\n\" world\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1561)\nA/1/2326\n\" life\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2326)\nA/1/2742\n\" violence\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2742)\nA/1/3084\nThis feature seems to fire on words ending in -ism, particularly when they are associated with schools of thought, philosophies, or belief systems.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3084)\nA/1/3498\nThis feature attends to the word \"his\", \"their\", \"path\", and similar words/concepts related to goals, opportunities, and progress.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3498)\nA/1/3614\n\" shame\", \" fear\", emotions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3614)\nCluster #281\nA/0/308\nThis feature attends to negative or insulting words that end in -ed or -er.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-308)\nA/1/470\nProfanity?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-470)\nA/1/3963\nMental health / wellness?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3963)\nA/1/3774\nTokens/nouns, esp. a few tokens after \"how\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3774)\nA/1/1778\nAbstract nouns after prepositions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1778)\nA/1/268\nAbstract verbs, ending in -s, related to imparting emotions or reputations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-268)\nA/1/3909\nEnds of words, mostly plural nouns (ending in -s), predicting ends of sentences?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3909)\nCluster #349\nA/0/320\nThe feature attends to plural nouns ending in -s or -es.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-320)\nA/0/449\nThis feature fires on words related to professions and groups of people.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-449)\nA/1/2391\nThe feature fires for words associated with professional roles such as researchers, analysts, curators, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2391)\nA/1/2762\nHuman roles/occupations in society\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2762)\nA/1/2766\nSocial/political identities\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2766)\nA/1/2929\nNumbers that increase by factors of 10\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2929)\nCluster #337\nA/1/303\nPotential final words in quotation, predicting punctuation with close-quote\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-303)\nA/1/1189\nThe feature fires when plural nouns particularly related to people end in \"s\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1189)\nA/1/4033\n\" people\"/\" children\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4033)\nA/1/2868\n\" men\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2868)\nA/1/2376\nEnd of quotations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2376)\nA/1/2675\n\" law\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2675)\nA/1/2983\n\" evidence\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2983)\nA/1/3385\n'peace' in diplomacy/news\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3385)\nA/1/3381\n\" Court\", predicting verbs describing court rulings like \" erred\" and \" remanded\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3381)\nCluster #149\nA/0/246\nThis feature seems to fire on legal or court related language, particularly around the concepts of judgments, appeals, jurisdiction and evidence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-246)\nA/1/1191\nThis feature seems to mainly fire on the phrase \"the validity\". It also often fires on phrases related to legal documents and court proceedings.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1191)\nA/1/1297\nAbstract legal nouns related to \"duty\"/\"defense\"/etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1297)\nA/1/1351\n'trial' before court in legal briefs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1351)\nA/1/2523\n\" judgment\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2523)\nA/1/3178\nThis feature fires when it sees legal language related to facts, such as \"facts and circumstances\", \"stipulated facts\", \"agreed based on facts\", \"need the facts\", \"proof\", \"requirements\", \"showing\", \"representations\", \"assertions\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3178)\nA/1/3739\nThe feature fires on sentences mentioning legal topics and terms related to contracts, protection, rights, court proceedings, evidence, compensation, employment, discharge, review, zoning, statutes, oaths, identity, complaints, remediation, negotiation, appeal, resignation and unions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3739)\nA/1/2728\nLegal terms like \" plaintiff\", \" defendant\", \" petitioner\" referring to parties in a legal case, esp. in legal principles and case law\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2728)\nCluster #12\nA/0/85\nThis feature fires for attention-grabbing capitalized content, such as quotes, slogans, or exclamations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-85)\nA/1/564\nShort quoted text inlined into longer sentences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-564)\nA/1/1526\nText shortly after \"smart quotes\" (U+201C)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1526)\nA/1/2571\nA first-initial-only capitalized token when words cannot tokenize with a prefixed space, as is often the case when beginning a quoted text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2571)\nA/1/2655\nAbstract nouns largely in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2655)\nCluster #156\nA/0/300\nThis feature fires when there is legal or economic terminology regarding contracts, liability, and related concepts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-300)\nA/1/955\nFinancial/legal text?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-955)\nA/1/1340\nLegal (Civil Law, Lawsuit, Torts)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1340)\nA/1/2309\nSentences relating to job employment\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2309)\nA/1/2471\nUS Constitutional Law\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2471)\nA/1/3647\nAbstract adjectives/verbs in credit/debt legal text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3647)\nA/1/3455\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3455)\nA/1/1167\nThis feature attends to the word \"unless\" and other similar words like \"although\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1167)\nCluster #144\nA/0/59\nThe feature fires when legal offenses and related terminology are present such as \"charged\", \"convicted\", \"sentence\", and \"crime\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-59)\nA/1/911\nCriminal law / sentencing language\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-911)\nA/1/1757\nLegal/crime reports?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1757)\nA/1/3704\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3704)\nA/1/1500\n' filed' in legal\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1500)\nCluster #157\nA/0/248\nThis feature fires when the text contains legal phrases related to presenting claims, evidence, arguments, issues and rulings in court cases.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-248)\nA/1/1842\n\" examining\" and other gerunds describing abstract actions in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1842)\nA/1/2044\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2044)\nA/1/2073\nVerbs describing court rulings like \" stated\" or \" concluded\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2073)\nA/1/2843\n\" denied\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2843)\nA/1/2513\nJudicial opinion / legal/court text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2513)\nA/1/42\nLegal / tax code past tense verbs or verb endings (\"ed\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-42)\nA/1/1130\n\" by\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1130)\nA/1/4032\nTitle Case (Military - partially army-specific)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4032)\nCluster #271\nA/0/389\nThis feature seems to fire for words associated with military topics and concepts, such as ranks, units, operations, weapons, equipment, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-389)\nA/1/1498\nArmy\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1498)\nA/1/1583\nShips and seafaring, describing voyages, wrecks, vessel types and terminology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1583)\nA/1/2807\nAviation and aerospace topics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2807)\nA/1/1391\nCrime logs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1391)\nCluster #238\nA/0/280\nThe feature seems to fire for proper nouns, particularly names of people, films, songs, and other works of art/media.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-280)\nA/1/714\nConnector words in film/TV/shows\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-714)\nA/1/862\nMusic ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-862)\nA/1/1332\nMovies / TV / Fiction (esp. action, superhero)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1332)\nCluster #336\nA/0/244\nThis feature attends to proper nouns and places, especially when referring to political topics such as countries, wars, conflicts, governments, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-244)\nA/0/500\nThe feature appears to fire on syntactic constructions related to sponsorship or support of policy or legislation, especially by politicians.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-500)\nA/1/127\nMiddle East ? esp. prepositions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-127)\nA/1/139\nMiddle-east organizations (conflict-related)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-139)\nA/1/941\nAmerican political parties\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-941)\nA/1/2273\nUK elections, especially Brexit\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2273)\nA/1/2654\nPolitical entity prediction\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2654)\nA/1/3410\nMueller investigation / Russiagate / Political Scandal\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3410)\nCluster #339\nA/1/1215\n\" armed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1215)\nA/1/1253\n\" state\" (legal)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1253)\nA/1/1595\nNationality/ethnicity adjectives, or ends thereof\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1595)\nA/1/2735\n\" former\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2735)\nA/1/3004\n\" police\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3004)\nA/1/3493\n\" military\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3493)\nCluster #11\nA/0/404\nThis feature fires for dialogue and quotes, attending to punctuation like quotation marks, as well as discourse markers like \"well,\" \"hey,\" and \"yes\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-404)\nA/1/143\n\" a\" and other articles/pronouns in quotes (dialogue)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-143)\nA/1/514\nNegated (after \"n't\" or \"not\") verbs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-514)\nA/1/730\nEnd of a name in quotes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-730)\nA/1/1094\nAcknowledgement of another character's point as a sentence fragment at the start of quoted speech\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1094)\nA/1/1655\nEnd of short clause prediction (in dialogue?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1655)\nA/1/2190\nVarious verbs esp. gerunds in quoted text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2190)\nA/1/2233\nNoun phrases (that are not pronouns) inside quotations, often after \"a\"/\"an\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2233)\nA/1/2970\nDirect objects, esp. pronouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2970)\nA/1/3858\nText in quotes, esp. imperative verbs / commands\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3858)\nA/1/2446\nVisual art\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2446)\nA/1/2506\nPolitical language, predicting a political ideology like \" socialism\"/\" democracy\"/\" capitalism\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2506)\nA/1/3326\nNarratives / conversations ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3326)\nA/1/102\nEnds of names or phrases specifying/near people; broadly fiction-plot-related\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-102)\nA/1/119\nPrepositions describing physical/locational relations between people (characters in fiction?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-119)\nCluster #226\nA/0/447\nThis feature fires when it sees words commonly associated with elevated/formal language and old-fashioned styles, such as older literature. It attends to words like \"thou\", \"hast\", \"thy\", \"verily\", archaic verb forms (\"knowest\", \"fathered\"), and titles like \"lord\" and \"sir\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-447)\nA/1/1639\nFlowery language?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1639)\nA/1/2103\nDescriptive language and old fashioned vocabulary?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2103)\nA/1/3335\nEarly Modern (Shakespearean) English\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3335)\nA/1/3925\nEnglish text after mid-sentence linebreaks\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3925)\nCluster #227\nA/0/356\nThe feature fires when the sentence contains prepositions like \"for\", \"in\", \"to\", \"with\", \"at\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-356)\nA/1/2577\nConjunctions / connectors like \"then\" and \"yet\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2577)\nA/1/2647\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2647)\nA/1/2894\nConstructs from excerpts in stories or personal narratives\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2894)\nA/1/3548\nVerbs and prepositions, esp. receiving people as the object, predicting female pronouns/titles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3548)\nA/1/26\nPersonal pronouns (\" I\"/\" you\"/...) in Biblical contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-26)\nCluster #228\nA/1/1042\nShort words immediately after em dashes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1042)\nA/1/1298\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1298)\nA/1/1309\nText of informal speech, inside speech marks \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1309)\nA/1/2396\nWords in political/religious/military text/speeches\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2396)\nA/1/2614\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2614)\nA/1/3223\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3223)\nA/1/3986\nEither token in \" ii.\" or nearby punctuation/numbers, typically volume references\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3986)\nA/1/163\n\"worship\" and other words/phrases of judeo christian texts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-163)\nA/1/3124\nMale pronouns, esp. \" himself\", or nearby words, predicting similar\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3124)\nA/0/147\nThe feature attends to words or phrases describing violence or destruction.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-147)\nCluster #320\nA/1/1815\nAdjectives for specific religions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1815)\nA/1/1824\n\"era\" concepts, both ancient greek/roman, medieval, and modern\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1824)\nA/1/3591\nHarmful/bad adjectives relating to prejudice and aggression\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3591)\nA/1/420\n\"ful\" and other word/word components esp. emotion-related\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-420)\nA/1/841\nAbstract, esp philosophical, -ical/-ive adjectives or endings, predicting further suffixes like -ism\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-841)\nCluster #241\nA/0/201\nThis feature attends to proper nouns and titles.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-201)\nA/0/373\nThis feature appears to activate when: 1. The text contains words ending in \"-tion\" or \"-ing\" 2. Prepositions like \"of\", \"to\", \"between\", etc. 3. Words associated with philosophical or academic writing (\"implications\", \"formulations\", \"predicated\", etc.) So in summary, this feature seems to fire on text with a philosophical/academic style, particularly words with the suffixes \"-tion\"/-\"ing\" and common prepositions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-373)\nA/0/448\nThis feature seems to attend to formal or archaic-sounding religious language, with a focus on worshipping, professing belief, preaching, praying, invoking God, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-448)\nA/1/49\nHistorical empires\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-49)\nA/1/179\nReligious (Christian)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-179)\nA/1/1701\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1701)\nA/1/1990\nPhilosophy\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1990)\nA/1/2788\nLinguistics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2788)\nA/1/2958\nSpiritual or religious texts about meditation, enlightenment and transcendence\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2958)\nA/1/3977\nPrepositions in contexts discussing poetry/poets\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3977)\nCluster #319\nA/0/165\nThis feature attends to the concept of politics and political structures.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-165)\nA/1/2427\n\" national\" and similar location-related terms, predicting suffixes like \"-ities\"/\"-ize\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2427)\nA/1/3215\n\" legislative\"/\" disciplinary\"/ some long words in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3215)\nA/1/3247\n\" public\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3247)\nA/1/3804\n\" foreign\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3804)\nA/1/1210\nThis feature fires on words that are generally considered negative but not strong negatives or curse words - words like \"unfair\", \"illegal\", \"dangerous\", \"sharp\", \"massive\", \"significant\". It does not activate on mild negatives like \"unnecessary\" or stronger negatives like \"terroristic\" or \"sacrificed\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1210)\nA/1/2550\n\" end\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2550)\nA/1/497\n\" top\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-497)\nA/1/2650\n\" middle\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2650)\nA/1/3477\n\"th\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3477)\nA/1/3105\nFirst/Second/Third/Next/Last prefixes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3105)\nA/1/4041\nCapitalized ordinal numbers (First, etc.) in legal contexts (predicting Amendment, Circuit, etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4041)\nCluster #369\nA/0/23\nThis feature seems to fire on ordinal numbers and words indicating something is first, next, last, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-23)\nA/1/269\n\" last\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-269)\nA/1/1317\nOrdinal numbers or sometimes fractions, e.g. \" third\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1317)\nA/1/2322\n\" final\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2322)\nA/1/3551\n\" instant\"/\"instant\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3551)\nA/1/3917\n\" hot\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3917)\nA/1/1165\n\" real\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1165)\nA/1/1781\n\" young\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1781)\nA/1/3913\nTechnique description, predicting a word for a technique\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3913)\nA/1/475\n\" open\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-475)\nA/1/3194\nAdjectives or ends thereof after \"a\"/\"an\", esp. \"thin\", describing material, thickness, or physical quality\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3194)\nA/1/1658\n\" common\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1658)\nCluster #305\nA/0/488\nThis feature attends to words and phrases related to pleasantness and politeness.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-488)\nA/1/104\nAdjectives/adverbs after \" more\"/\" less\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-104)\nA/1/202\nThis feature fires for the concept of greater magnitude, attending to words like \"stronger\", \"louder\", \"more evident\", \"more visible\", etc. It seems to capture the idea of something being bigger, faster, longer, more prominent, more apparent, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-202)\nA/1/313\nThis feature attends to concepts related to difficulty, complexity, or hesitance.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-313)\nA/1/612\n\" good\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-612)\nA/1/1169\n\" better\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1169)\nA/1/3030\n\" right\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3030)\nA/1/3243\n\" hard\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3243)\nA/1/3319\n\" fun\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3319)\nA/1/3364\nThis feature seems to fire on sentences describing something as amusing, funny or interesting.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3364)\nA/1/3464\n\" close\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3464)\nA/1/3791\n\" free\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3791)\nCluster #368\nA/0/27\nThe feature fires for sentences that describe something as pleasant or wonderful.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-27)\nA/0/143\nThe feature fires when it encounters words and phrases that are often used in formal or technical writing, such as \"furthermore\", \"structured partials\", \"known to benefit\", \"fully targeted lessons\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-143)\nA/1/64\n\" general\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-64)\nA/1/296\n\" new\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-296)\nA/1/499\n\" special\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-499)\nA/1/720\n\" past\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-720)\nA/1/733\n\" same\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-733)\nA/1/773\nSuperlatives like biggest/greatest/largest\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-773)\nA/1/1120\n'particular' as a token\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1120)\nA/1/1921\n\" recent\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1921)\nA/1/2067\n\" early\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2067)\nA/1/2338\nStrong-valence adjectives\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2338)\nA/1/2366\n\" prior\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2366)\nA/1/3065\n\" best\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3065)\nA/1/3145\n\" double\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3145)\nA/1/3507\n\" main\" / some synonyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3507)\nA/1/4037\n\" big\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4037)\nA/1/8\n\" fair\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-8)\nA/1/2698\n\" later\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2698)\nCluster #313\nA/1/252\n\" full\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-252)\nA/1/584\n\" wide\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-584)\nA/1/1103\n\" long\" token, low-activating \" short\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1103)\nA/1/1972\n\" small\"/\" large\"/size-related\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1972)\nA/1/1850\n\" upper\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1850)\nCluster #314\nA/0/3\nThis feature attends to words or phrases associated with low to high levels or ranges.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-3)\nA/1/340\nThis feature fires for words and phrases related to similarity and identicalness.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-340)\nA/1/727\n\" equal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-727)\nA/1/1143\n\" high\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1143)\nA/1/1593\n\" positive\"/\" negative\" in scientific/technical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1593)\nA/1/2286\n\" higher\"/\" lower\" comparing values\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2286)\nA/1/3481\n\" significant\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3481)\nA/1/3229\nThis feature attends to words indicating increase, decrease or other changes in quantities/measurements.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3229)\nA/1/1859\n\" inner\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1859)\nA/1/3311\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3311)\nA/1/2709\n\" little\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2709)\nA/1/1629\n\" maximum\"/\" minimum\"/etc. in vaguely scientific contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1629)\nA/1/1745\nThe feature attends to words with the adjective/past participle suffix \"-ed\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1745)\nCluster #334\nA/0/203\nThis feature seems to fire on the word \"one\" when used numerically, either for counting or as an ordinal number.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-203)\nA/1/1522\n\" one\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1522)\nA/1/2161\n\" one\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2161)\nA/1/2758\n\" every\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2758)\nA/1/2954\n\" One\"/\"One\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2954)\nA/1/3359\n\" another\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3359)\nA/1/3630\n\" half\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3630)\nA/1/3811\n\" other\" or \" another\" esp. after \" each\" or \" one\", though sometimes \" the\" after \" one[ of]\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3811)\nA/1/4055\n\" each\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4055)\nA/1/1222\nAdjectives in economic/financial news\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1222)\nA/1/932\nValenced adjectives after \" a\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-932)\nA/0/37\nThis feature seems to fire when there are words related to quantitative descriptions and measurements, especially broader or larger quantities. Some key words it attends to are \"substantial\", \"large\", \"significant\", \"broad\", \"additional\", \"strong\", \"increased\", \"practical\", \"direct\", \"dramatic\", \"array\", etc. Quantitative terms like percentages, time periods, ranges, and other measurements also tend to trigger higher activations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-37)\nA/1/762\n\" present\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-762)\nA/1/2610\n\" omn\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2610)\nA/1/756\n\" very\", \" rather\", adverbs indicating extent of something\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-756)\nA/1/2285\nPhysical adjectives, maybe in mechanical or material engineering/design\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2285)\nA/1/2971\n\" one\"/\" only\" after \"the\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2971)\nA/1/2783\nQualitative descriptors in scientific context, e.g. \" unique\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2783)\nA/1/1797\n\" exact\", \" specific\", and similar abstract adjectives\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1797)\nA/1/1597\n\" following\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1597)\nA/1/2773\n\" above\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2773)\nA/1/4085\nAdjectives about people esp. after possessive pronouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4085)\nA/1/2098\nThis feature fires for words that indicate transitions such as new ideas, new topics, new paragraph or sections. Examples include words like \"the latter\", \"following\", \"thus\", \"however\" and ordinal numbers like \"fourth\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2098)\nA/1/1459\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1459)\nA/1/1704\nPossessive pronouns: \"his\", \"its\", \"their\", in religious/theological contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1704)\nA/1/1395\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1395)\nA/1/2791\nChristian church management\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2791)\nCluster #244\nA/1/328\nPossessive pronouns in religious/mythical contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-328)\nA/1/1818\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1818)\nA/1/3301\n\" his\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3301)\nA/1/43\nLegal adjectives usu. describing abstract nouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-43)\nCluster #255\nA/0/342\nThis feature seems to fire when it detects words relating to courts and legal proceedings, such as \"trial\", \"court\", \"district court\", \"judge\", \"petitioner\", \"plaintiff\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-342)\nA/1/829\n\" the\" in police reports?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-829)\nA/1/1106\n\" the\", in legal context\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1106)\nA/1/1563\n'the' in legal\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1563)\nA/1/2184\n\" this\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2184)\nA/1/2718\n\" a\" in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2718)\nA/1/3878\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3878)\nA/1/213\n'his' as possessive prnoun in legal\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-213)\nCluster #245\nA/0/456\nThis feature seems to attend to possessive determiners like \"my\", \"his\", \"her\", and \"their\". Specifically it fires strongly when \"my\" is present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-456)\nA/1/41\n\" His\" (esp. familial contexts?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-41)\nA/1/348\n\" its\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-348)\nA/1/429\n\" our\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-429)\nA/1/711\n\" her\", predicting a person-relationship word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-711)\nA/1/967\n\" Her\"/\"Her\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-967)\nA/1/1090\nPossessive pronouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1090)\nA/1/2031\n\" My\"/\"My\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2031)\nA/1/2558\n\" his\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2558)\nA/1/2888\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2888)\nA/1/3099\n\" your\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3099)\nA/1/3303\n\" their\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3303)\nA/1/3422\n\" my\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3422)\nA/1/3813\n\" own\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3813)\nA/1/3833\nThis feature attends to words indicating possession such as \"my\", \"your\", \"his\", and \"their\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3833)\nCluster #254\nA/0/1\nThis feature attends to the word \"The\" at the beginning of sentences.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-1)\nA/0/70\nThis feature attends to the word \"the\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-70)\nA/0/78\nThe feature fires when it sees the word \"the\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-78)\nA/0/438\nThis feature fires when the phrase \"the\" is present in the example.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-438)\nA/0/510\nThis feature attends to the word \"the\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-510)\nA/1/1\n\"the\" in a project management context\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1)\nA/1/69\nUsually \" the\", after \" at\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-69)\nA/1/100\n\" the\" token in \"on the\", generic use\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-100)\nA/1/228\n\" the\" after \"all\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-228)\nA/1/474\n\" the\"/\"The\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-474)\nA/1/798\n\" the\" usually before a location\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-798)\nA/1/1039\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1039)\nA/1/1063\nDiscussions of books in general?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1063)\nA/1/1110\n'the' preceding <location>\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1110)\nA/1/1148\n\"The\"/\" The\" in book titles, esp. underscore-delimited\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1148)\nA/1/1293\n\" the\" in geography?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1293)\nA/1/1304\n'the' before a title\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1304)\nA/1/1674\n\" the\" in \"over the\" event or time period\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1674)\nA/1/1751\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1751)\nA/1/1898\n\" the\" in aeronautics? sea travel?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1898)\nA/1/2116\n\" the\" in discussion of music/film\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2116)\nA/1/2206\n\" the\" in royalty-related contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2206)\nA/1/2288\n\" the\" in a medical context\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2288)\nA/1/2811\n\" the\" in economics?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2811)\nA/1/2879\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2879)\nA/1/2916\n\" the\" in descriptive writing, esp physical descriptions of people?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2916)\nA/1/2931\n\" the\" in Biblical contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2931)\nA/1/2932\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2932)\nA/1/3199\n\" the\" in (U.S.) politics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3199)\nA/1/3358\n\" The\"/\"The\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3358)\nA/1/3378\n\" the\" in international relations / geopolitics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3378)\nA/1/3388\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3388)\nA/1/3407\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3407)\nA/1/3530\n\" the\" as objects of verbs, possibly military/political/physical contexts?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3530)\nA/1/3564\n\" the\" in philosophy/psychology?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3564)\nA/1/3685\n\"the\" in Fantasy games\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3685)\nA/1/3937\n\" the\" in formal English?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3937)\nCluster #256\nA/0/220\nThis feature attends to the word \"the\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-220)\nA/1/5\n\"the\" in electrical engineering contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-5)\nA/1/676\n\" the\" in photography? optics?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-676)\nA/1/731\n\" the\" in mechanical engineering\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-731)\nA/1/1746\n\" the\" in chemistry\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1746)\nA/1/2799\n\" a\" in mechanical engineering\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2799)\nCluster #249\nA/1/234\n\" a\" esp. after prepositions like \"by\"/\"with\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-234)\nA/1/2068\n\" a\" after \"with\" and other prepositions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2068)\nA/1/2942\n\" the\" after \"is\"/\"be\"... predicting superlatives?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2942)\nA/1/3104\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3104)\nCluster #250\nA/0/34\nThis feature fires when the word \"a\" appears as an indefinite article in sentences. It gives especially high activation to sentences where \"a\" is followed by a noun, but also activates for other uses of \"a\" as an indefinite article, such as before adjectives.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-34)\nA/0/417\nThe feature seems to fire most strongly when the word \"a\" is used immediately before a descriptive noun phrase.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-417)\nA/1/230\n\" a\"/\" an\" after \" as\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-230)\nA/1/674\n\" a\" esp. after \"make\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-674)\nA/1/1727\n\" a\" after \"in\", sometimes after other prepositions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1727)\nA/1/2024\n\" a\" after \"has\" or sometimes \"is\" and word forms or contractions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2024)\nA/1/2264\n\" a\" after \"with\" or some gerunds\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2264)\nA/1/2434\n\" a\" esp. after names / proper nouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2434)\nA/1/3516\n\" a\" after \"is\"/etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3516)\nA/1/3576\n\" a\" after \"take\", \"get\", \"give\", etc. and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3576)\nA/1/3766\n\" a\" in political or current events contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3766)\nA/1/4078\n\" a\", sometimes after \"like\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4078)\nCluster #248\nA/0/484\nThis feature fires when the word \"an\" appears before a word starting with a vowel sound.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-484)\nA/1/666\n\" An\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-666)\nA/1/3635\n\" an\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3635)\nA/1/2384\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2384)\nCluster #247\nA/0/439\nThe feature seems to fire whenever it sees the determiner \"this\" being used in a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-439)\nA/1/935\n\" this\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-935)\nA/1/952\n\" This\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-952)\nA/1/1281\n\" this\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1281)\nA/1/1584\n\" This\"/\"This\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1584)\nA/1/3758\n\" this\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3758)\nCluster #246\nA/0/74\nThis feature seems to respond most strongly to the word \"the\", especially when it is being used in a grammatical sense (e.g. \"the car\", \"the problem\", \"the server\" etc). It also occasionally fires for other determiners/articles like \"a\" and \"an\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-74)\nA/1/820\n\" a\" as object of verb or preposition in programming contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-820)\nA/1/1175\nCode: \" the\" in code comments?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1175)\nA/1/1400\n\" my\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1400)\nA/1/3893\n\" an\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3893)\nA/1/3054\nThis feature attends to concepts related to possession, firing primarily on tokens such as \"their\", \"his\", \"its\", \"our\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3054)\nCluster #257\nA/0/90\nThis feature attends to the definite article \"the\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-90)\nA/1/1214\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1214)\nA/1/1520\n\" The\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1520)\nA/1/1857\n\" a\" esp. after \"showed\" or synonyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1857)\nA/1/1893\n\" the\" in medical/anatomical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1893)\nA/1/2400\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2400)\nA/1/2841\n\" the\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2841)\nA/1/3784\n\" the\" in biology/pathology?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3784)\nCluster #56\nA/0/490\nThe feature fires when there are instances of the word \"no\" in the text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-490)\nA/1/1101\n\"Any\" token, w or w/o leading space\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1101)\nA/1/1622\n\" no\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1622)\nA/1/1789\n\" any\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1789)\nA/1/2033\n\" no\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2033)\nA/1/3089\n\" No\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3089)\nA/1/2526\n\" These\" in science?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2526)\nCluster #55\nA/0/495\nThis feature attends to descriptive words that indicate quantity, such as numbers, \"some\", \"any\", \"several\", \"many\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-495)\nA/1/142\n\" those\" single-token feature\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-142)\nA/1/291\nSmall spelled-out English numbers, esp. in scientific contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-291)\nA/1/807\n\" multiple\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-807)\nA/1/937\n\" many\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-937)\nA/1/1470\n\" different\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1470)\nA/1/2006\nSmall spelled-out English numbers, esp. measuring units of time\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2006)\nA/1/2529\n\" These\"/\" these\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2529)\nA/1/2587\nLarge numbers written out in words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2587)\nA/1/3116\nAdjectives after words stating number or quantity like \"two\" or \"several\", especially suggesting comparison\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3116)\nA/1/3234\n\" other\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3234)\nA/1/3978\n\" some\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3978)\nCluster #53\nA/0/264\nThe feature seems to fire on the word \"all\" in most cases. More specifically, it appears to activate when \"all\" is used as an intensifier to emphasize something, like \"all the books\" or \"giving everything up\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-264)\nA/1/1394\nThis feature seems to fire when it encounters words related to duality or a choice between two options, such as \"both,\" \"either,\" \"neither,\" \"outside,\" etc. It appears sensitive to expressions involving two alternatives or opposites.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1394)\nA/1/1623\n\" All\"/\" all\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1623)\nA/1/1868\n\" none\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1868)\nA/1/922\n's token in legal\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-922)\nCluster #50\nA/0/453\nThis feature seems to fire when it encounters possessive nouns and pronouns ending in 's.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-453)\nA/1/1108\n\"s\" after \\u2019 unicode apostrophe\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1108)\nA/1/1545\n\"'s\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1545)\nA/1/3583\nThe feature fires when it sees possessive nouns, particularly proper nouns ending in 's.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3583)\nA/1/3710\n\"'s\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3710)\nCluster #51\nA/1/611\n\"'s\" in contracted pronouns: he's, she's and occasionally who's and nobody's\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-611)\nA/1/1714\n\"'s\" in \"That's\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1714)\nA/1/3183\n\"It's\" __\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3183)\nA/1/3001\n\"'s\" in \"There's\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3001)\nA/1/3242\n\"m\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3242)\nCluster #44\nA/0/219\nThis feature fires on apostrophes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-219)\nA/1/792\n\"\\u2019\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-792)\nA/1/1176\n\"\\u2019\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1176)\nA/1/2492\n\"\\u2019\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2492)\nA/1/3420\n\"\\u2019\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3420)\nA/1/3580\n\"'\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3580)\nA/1/3657\n\"\\u2019\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3657)\nA/1/4094\nU+2019, apostrophe, after \"It\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4094)\nA/1/2148\n\"re\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2148)\nA/1/1735\n\" AL\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1735)\nA/1/1266\nALLCAPS fragments ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1266)\nCluster #125\nA/0/317\nThis feature fires on all uppercase words or short phrases.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-317)\nA/0/421\nThe feature fires for words written in all caps.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-421)\nA/1/704\n\"EE\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-704)\nA/1/791\nUppercase formal Greek Unicode character names? (between \"{~\" \"~}\"?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-791)\nA/1/1100\nALL CAPS, usually at start of line\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1100)\nA/1/1150\n\" AN\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1150)\nA/1/1348\n\" THE\"/\"THE\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1348)\nA/1/1359\n\" CON\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1359)\nA/1/1527\nER or SON name ending when referring to judges\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1527)\nA/1/1556\nAll-caps word beginnings esp. two letters ending in O\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1556)\nA/1/1633\n\" IN\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1633)\nA/1/1667\n\" AND\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1667)\nA/1/1812\nUPPERCASE Suffix / End of word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1812)\nA/1/1980\n\" RO\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1980)\nA/1/2011\nAll-caps beginnings of words (or, more rarely, abbreviations), particularly \" NE\" (predicting \"VER\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2011)\nA/1/2018\n\" SE\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2018)\nA/1/2398\n\"RE\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2398)\nA/1/2606\n\" IS\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2606)\nA/1/2631\n\" LI\"/\"LI\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2631)\nA/1/2810\n\" DO\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2810)\nA/1/2821\nThis feature fires when it sees the word announcement or variations like announce or announced. This applies across different contexts from formal press releases to informal social media posts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2821)\nA/1/2926\n\" MA\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2926)\nA/1/2956\n\" HE\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2956)\nA/1/3094\nThe feature fires for an acronym of 3-4 capital letters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3094)\nA/1/3197\nAll-caps names near phone numbers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3197)\nA/1/3269\nALL CAPS TEXT, esp. in screenplays/dialogue\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3269)\nA/1/3405\nAll-capital legal documents about contracts and public records\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3405)\nA/1/3439\n\" DI\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3439)\nA/1/3565\nCapital letter prefix that seems to predict the rest of the word will be in all caps\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3565)\nA/1/3753\n\"UR\" and other mid-word all caps English\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3753)\nA/1/3782\n\" EX\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3782)\nA/1/3801\n\" COMP\"/\"COMP\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3801)\nA/1/2976\n\" CO\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2976)\nA/1/2299\n\" CH\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2299)\nCluster #126\nA/0/412\nThis feature attends to acronyms and abbreviated terms, as it fires strongly on uppercase sequences of 2-4 characters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-412)\nA/1/527\nRoman numerals\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-527)\nA/1/2657\nThe feature fires when it sees acronyms or abbreviations in all caps.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2657)\nCluster #108\nA/0/136\nThis feature appears to fire on scientific acronyms, particularly those related to biology and chemistry.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-136)\nA/1/582\nShort ALLCAPS fragments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-582)\nA/1/1496\n\" SC\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1496)\nA/1/1712\nAstronomy and physics acronyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1712)\nA/1/1953\nCapital letters in abbreviations, esp. when not the beginning of the abbreviation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1953)\nA/1/2672\nAll-caps letters in abbreviations, esp. tokens after the first (without a preceding space)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2672)\nCluster #119\nA/1/18\n\" ST\" with other uppercase letter sequences lower\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-18)\nA/1/1029\nLast letter in all-caps abbreviations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1029)\nA/1/2739\n\"AT\" in ATM and AT&T acronyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2739)\nA/0/503\nThe feature seems to be activated by the presence of single capital letters, especially R and T, in technical contexts. Often these appear as abbreviations in formulas or chemical names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-503)\nA/1/2528\nThis feature attends to abbreviations related to chemistry, biology and medicine.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2528)\nA/1/1663\nShort letter sequences esp. ALLCAPS in technical contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1663)\nA/1/2974\nFrom the examples above, it seems like this feature attends to acronyms and initialisms, particularly three-letter acronyms/initialisms in all caps.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2974)\nCluster #87\nA/1/320\nALLCAPS abbreviation fragments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-320)\nA/1/1959\nBeginnings of all-caps abbreviations/initialisms, particularly predicting \"MA\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1959)\nA/1/2072\nSpace+capital-letter fragments usu. in abbreviations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2072)\nCluster #79\nA/0/33\nThis feature attends to capital letters like M, L, E, G, R and T in the middle of words or acronyms.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-33)\nA/1/45\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-45)\nA/1/547\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-547)\nA/1/1987\nSpace+letter tokens esp. in technical writing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1987)\nA/1/2344\nThis feature fires for proper nouns that start with the letters \"M\" or \"N\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2344)\nA/1/2498\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2498)\nA/1/3066\n\" Q\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3066)\nA/1/3604\nSpace + capital letter tokens, esp. \" Q\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3604)\nCluster #74\nA/0/344\nThis feature attends to acronyms and abbreviations for biological and chemical concepts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-344)\nA/0/489\nThis feature attends to DNA sequences with codes like CTT, GAG, AAT, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-489)\nA/1/68\n\" CD\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-68)\nA/1/161\nALLCAPS abbreviations in biology?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-161)\nA/1/830\nCapital letters beginning biology/chemistry abbreviations for molecules etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-830)\nA/1/875\nScientific nomenclature for genes and proteins\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-875)\nA/1/1353\nDNA/RNA, fires especially on directionality label\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1353)\nA/1/2151\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2151)\nA/1/2435\nThe feature is attending to protein abbreviations such as SHP, S1P, TLR4, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2435)\nA/1/2937\nDNA (upper case)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2937)\nA/1/2982\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2982)\nA/1/3027\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3027)\nA/1/3173\n\" IL\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3173)\nA/1/3392\nLowercase letter fragments in medical abbreviations, e.g. proteins or genes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3392)\nA/1/3412\nVarious ALLCAPS fragments predicting \"K\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3412)\nA/1/2281\nDNA (lower case)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2281)\nA/1/1944\nThis feature attends to the protein p and numbers that appear to be p-values from scientific experiments.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1944)\nCluster #112\nA/0/348\nThis feature fires when it sees italicized text or italicized abbreviations/symbols in scientific writing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-348)\nA/1/507\nText, particularly scientific names of organisms, surrounded/italicized by *'s\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-507)\nA/1/1856\nLetters in Molecules \"Eg O2\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1856)\nA/1/3017\nMostly-small physical dimensions: mm/cm distances for diameter/thickness/height/length\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3017)\nA/1/3146\n\"P\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3146)\nA/1/3588\n\" mg\", \" μg\", unit abbreviations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3588)\nA/1/4053\nC representing Celsius \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4053)\nA/0/181\nThis feature appears to activate for measurement symbols and abbreviations, such as m for meters/mass, mL for milliliters, mm for millimeters, s for seconds, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-181)\nCluster #113\nA/0/374\nThis feature fires when detecting certain protein, genetic, and biochemical concepts such as mentions of specific genes, enzymes, and processes like phosphorylation. It attends to both gene symbols (e.g. CRE1, RCK1) as well as written out names of proteins and cell types (caspase-3, CD25+ T regulator cells). Numbers that indicate genetic loci and molecular weights also activate the feature.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-374)\nA/1/560\nNumbers esp. after capital letters (biology?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-560)\nA/1/1310\nNumber in CDn genes, eg. 57 in CD57\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1310)\nA/1/2656\nThis feature fires when it sees numbers with a single digit (1-9).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2656)\nA/1/2256\n\" most\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2256)\nCluster #273\nA/0/371\nThe feature fires on the word \"more\" and words related to degrees and comparisons such as \"most\", \"equally\", \"highly\", etc. It seems to be attending to words used for comparisons.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-371)\nA/1/199\n\" more\" and some \" less\"/\" most\" after \" is\"/\" are\"/\" was\"/etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-199)\nA/1/971\n\" more\"/\" less\" in scientific contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-971)\nA/1/1077\n\" much\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1077)\nA/1/1462\n\" More\"/\"More\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1462)\nA/1/1999\n\" more\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1999)\nA/1/2625\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2625)\nCluster #284\nA/0/458\nThis feature attends to the use of adverbs that end in \"-ly\" in the text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-458)\nA/1/329\n\" highly\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-329)\nA/1/649\nThis feature fires for words related to doing something automatically or by hand, such as \"manually\", \"automatically\", \"directly\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-649)\nA/1/901\n\" closely\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-901)\nA/1/1062\n\" rather\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1062)\nA/1/1213\nThe feature appears to fire on words with the -ically suffix, which denotes \"in a manner of\" or \"pertaining to\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1213)\nA/1/2592\n\" completely\"/\" totally\" adverbs indicating extremeness\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2592)\nA/1/3278\n\" well\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3278)\nA/1/3650\nAdverbs indicating short amounts of time: \" soon\"/\" shortly\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3650)\nA/1/3765\n\" far\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3765)\nA/1/4000\nThis feature fires on adverbs that end in -ly.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4000)\nA/1/2319\n\" further\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2319)\nA/1/405\n\" simply\" and \" always\" after \"is\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-405)\nCluster #278\nA/0/464\nThis feature fires when it detects words related to emphasis and intensity such as \"very\", \"quite\", \"totally\", \"extremely\", \"absolutely\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-464)\nA/1/67\n\" very\" with some other intensifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-67)\nA/1/1572\n\" relatively\" predicting size/money\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1572)\nA/1/1874\n\" too\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1874)\nA/1/2318\n\" pretty\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2318)\nCluster #262\nA/0/483\nThe feature appears to activate when a verb is used to describe something that typically or frequently occurs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-483)\nA/1/758\n\" also\" and similar frequency modifiers in medicine/science?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-758)\nA/1/2984\n\" significantly\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2984)\nA/1/3374\n\" nearly\"/\" almost\"/synonyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3374)\nA/1/2838\n\"seem\" \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2838)\nCluster #283\nA/0/415\nThis feature seems to fire when the word \"just\" is used.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-415)\nA/0/451\nThis feature seems to fire when the word \"so\" is used as a conjunction or adverb to emphasize or indicate consequence. It attends to sentences where \"so\" connects clauses or comes before adjectives and adverbs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-451)\nA/1/2\n\"also\" after an institution\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2)\nA/1/749\n\" just\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-749)\nA/1/1041\n\" particularly\" and some synonyms, esp. after \",\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1041)\nA/1/2449\n\" once\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2449)\nA/1/2491\nThis feature attends to formal or descriptive words that indicate typicality or commonality.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2491)\nA/1/2839\nLocalizing events temporally - the relative timing of events as ongoing (\"still\"), completed ahead of schedule (\"already\"), or concurrent (\"currently\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2839)\nA/1/3383\n\" exactly\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3383)\nA/1/3594\n\" really\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3594)\nA/1/3696\n\" even\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3696)\nA/1/1404\n\" never\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1404)\nA/1/2053\n\" so\", sometimes after contracted \"is\"/\"are\" or dashes/hyphens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2053)\nA/1/3903\n\" only\" token\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3903)\nA/1/3519\n\" now\" and other time-related words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3519)\nA/1/2074\n\" always\" / intensifiers or frequency-indicating adverbs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2074)\nA/0/134\nThis feature fires on phrases that express negation, uncertainty, impossibility, or doubt. Some examples include \"cannot\", \"not\", \"never\", \"doesn't\", \"unlikely\", etc. It seems to activate most strongly on the words \"not\" and \"cannot\", but also fires on other negative phrases as well.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-134)\nA/1/3919\n\" Not\"/\"Not\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3919)\nA/1/2433\n\" not\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2433)\nA/1/1753\n\"'t\" in contractions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1753)\nA/1/389\n\" not\" after \"does\" and similar\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-389)\nA/1/974\nAuxiliary verbs (in technical writing/docs?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-974)\nCluster #268\nA/0/252\nThis feature attends to contractions with n't.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-252)\nA/1/169\n\"t\", usually after U+2019 as apostrophe in contraction\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-169)\nA/1/770\n\"'t\" in \" couldn't\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-770)\nA/1/3600\n\"'t\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3600)\nA/1/317\n\"Can\"/\" Can\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-317)\nA/1/2830\n\"ve\" in \"I’ve\" / other \" 've\" constructions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2830)\nCluster #267\nA/0/117\nThis feature attends to modal verbs like \"would,\" \"could,\" and \"might.\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-117)\nA/0/142\nThis feature fires when the text includes words like \"could\", \"may\", \"should\", \"might\", \"would\", etc. that indicate possibility, uncertainty, or conditional statements.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-142)\nA/0/215\nThe feature attends to sentences where the modal verb \"can\" is used.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-215)\nA/1/350\n\" may\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-350)\nA/1/456\n\" can\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-456)\nA/1/610\n\" can\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-610)\nA/1/780\n\" shall\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-780)\nA/1/1192\n\" would\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1192)\nA/1/2536\n\" simply\" and other adverbs after \"can\"/\"could\"/\"would\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2536)\nA/1/2804\n\" can\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2804)\nA/1/3212\n\" must\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3212)\nA/1/3534\n\" will\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3534)\nA/1/3835\n\"'ll\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3835)\nA/1/2380\n\"'d\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2380)\nA/1/3080\n\" Is\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3080)\nA/1/338\n\" Do\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-338)\nCluster #239\nA/0/391\nThe feature is attending to past and present verb tenses, particularly those using the word \"has\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-391)\nA/1/256\n\" has\"/\" have\" usu. after singular nouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-256)\nA/1/441\n\" have\" after \" would\"/\" might\"/...\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-441)\nA/1/2371\n\" has\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2371)\nA/1/2372\n\" has\"/\" have\" in scientific contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2372)\nA/1/2693\n\" Has\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2693)\nA/1/3156\n\" have\" directly after plurals, usu. groups of people\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3156)\nA/1/3216\n\" had\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3216)\nA/1/3308\n\" has\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3308)\nA/1/3659\n\" have\" after \"I\"/\"You\" in double quotes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3659)\nA/1/3823\n\" have\"/\" having\" in technical contexts, asking questions or troubleshooting\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3823)\nA/1/2292\nObject pronouns after \"let\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2292)\nCluster #235\nA/0/107\nThis feature fires on present tense conjugated forms of the verb \"to be\" in negative contexts such as don't, isn't, weren't, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-107)\nA/0/121\nThis feature attends to contractions and short forms of the words \"do not\" and \"does not\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-121)\nA/1/579\n\" won\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-579)\nA/1/1472\n\" didn\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1472)\nA/1/1577\n\" don\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1577)\nA/1/2174\n\" Don\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2174)\nA/1/2497\n\" isn\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2497)\nA/1/3928\nThe feature seems to fire most strongly when the word \"did\" appears. This suggests it is particularly attuned to past tense verb phrases. The feature does not appear to be sensitive to any particular semantic content or style, simply responding most strongly to sentences with past tense verbs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3928)\nCluster #195\nA/0/125\nThis feature fires for sentences containing the phrase \"like\" or \"about\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-125)\nA/1/1770\n\" like\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1770)\nA/1/2948\n\" about\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2948)\nA/1/3175\n\" with\" describing agreement/opinions/responsibility\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3175)\nA/1/3851\n\" like\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3851)\nCluster #180\nA/0/28\nThis feature fires when a sentence starts with a clause that ends with the word \"after\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-28)\nA/1/180\n\" after\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-180)\nA/1/187\n\" during\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-187)\nA/1/3149\n\" before\", predicting \"-ing\" participles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3149)\nA/1/933\nThis feature seems to fire when it encounters the preposition \"among\" connecting different nouns or noun phrases.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-933)\nCluster #221\nA/0/7\nThe feature seems to fire when the word \"by\" appears in the middle of sentences. It appears to detect the preposition \"by\" that connects a verb phrase to its object.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-7)\nA/0/273\nThe feature seems to fire when the word \"with\" is used in a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-273)\nA/1/71\n\" for\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-71)\nA/1/487\n\" for\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-487)\nA/1/857\n\" include\" and word forms, usually introducing examples\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-857)\nA/1/2379\n\" without\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2379)\nA/1/2580\n\" by\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2580)\nA/1/3416\n\" with\" in finance/economics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3416)\nA/1/3524\n\" with\" in stories? folktales?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3524)\nA/1/1866\n\" such\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1866)\nA/1/1847\n\" between\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1847)\nA/1/2589\n\" By\"/\"By\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2589)\nCluster #165\nA/0/42\nBased on the examples, this feature seems to fire most strongly when it detects the phrase \"one of the\" or variations of it (e.g. \"one of\", \"one of these\", etc). This indicates that it may be attending to this common phrase structure.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-42)\nA/0/328\nThe feature fires for the phrases \"of the\", \"of\", and \"University of\" which are commonly found in organization names. So it seems to be paying attention to organization names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-328)\nA/0/419\nThe feature fires when it sees the preposition \"of\" followed by a noun phrase.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-419)\nA/0/428\nThis feature fires when it sees the preposition \"of\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-428)\nA/1/243\n\" of\" in \"Board of ...\" and similar administrative units of something\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-243)\nA/1/314\nThe feature fires when the word \"of\" is used to indicate something is comprised or consists of something else.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-314)\nA/1/343\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-343)\nA/1/349\n\" of\" after noun, esp. food/drink/clothing\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-349)\nA/1/407\n\" of\" in a legal document\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-407)\nA/1/492\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-492)\nA/1/786\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-786)\nA/1/916\n\" Of\"/\"Of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-916)\nA/1/1067\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1067)\nA/1/1128\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1128)\nA/1/1543\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1543)\nA/1/1553\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1553)\nA/1/2061\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2061)\nA/1/2181\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2181)\nA/1/2198\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2198)\nA/1/2227\n\"of\" in university names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2227)\nA/1/2334\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2334)\nA/1/2628\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2628)\nA/1/2691\n\"of the\" in patent statement\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2691)\nA/1/2694\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2694)\nA/1/2826\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2826)\nA/1/2832\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2832)\nA/1/2854\n\"of\" in History Book Titles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2854)\nA/1/3563\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3563)\nA/1/3718\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3718)\nA/1/3840\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3840)\nA/1/3900\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3900)\nA/1/664\n\" for\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-664)\nA/0/187\nThe feature fires when it sees the word \"for\" in contexts where it means \"because.\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-187)\nCluster #166\nA/0/13\nThe feature seems to fire when the word \"at\" is used to indicate a location or specific place.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-13)\nA/1/284\n\" at\" in legal contexts referencing a (page) number?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-284)\nA/1/620\nThis feature fires for words that indicate a time or moment.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-620)\nA/1/680\n\" at\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-680)\nA/1/1808\n\" At\"/\"At\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1808)\nA/1/1939\n\" at\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1939)\nA/1/3125\n\" at\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3125)\nA/1/3665\n\" at\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3665)\nA/1/1958\n\" On\"/\"On\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1958)\nA/1/2382\n\" For\"/\"For\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2382)\nA/1/1867\n\" under\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1867)\nA/1/898\nPrepositions about (U.S.?) historical figures/events, predicting 18xx years?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-898)\nCluster #222\nA/0/71\nThis feature attends to the word or phrase \"over\" used as a preposition in a variety of contexts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-71)\nA/0/253\nThis feature seems to fire when there are words related to moving or transitioning from one thing to another. Examples include \"from\", \"into\", \"towards\", \"away\", \"through\", etc. It detects words and phrases that indicate a change in position, location, or state.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-253)\nA/0/441\nThis feature attends to the preposition \"on\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-441)\nA/1/494\n\" around\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-494)\nA/1/1038\n\" from\", often in government or geography contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1038)\nA/1/1182\n\" along\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1182)\nA/1/1230\n'on' token, generic use\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1230)\nA/1/1292\n\" over\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1292)\nA/1/1315\n'on' meaning 'in relation to', eg. in 'based on'\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1315)\nA/1/1409\n\" beyond\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1409)\nA/1/1488\n\"across\" a spatial/geographical area\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1488)\nA/1/1644\n\" through\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1644)\nA/1/1681\n'on' in technical text, mostly referring to information contained in some reference\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1681)\nA/1/1894\n\" under\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1894)\nA/1/2028\n\" into\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2028)\nA/1/2059\n\" Over\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2059)\nA/1/2378\n\" against\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2378)\nA/1/3402\n\" within\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3402)\nCluster #171\nA/0/115\nThis feature fires when the phrase \"in\" appears by itself or as part of another word or phrase. It often activates in sentences where \"in\" is used to indicate location or being part of something, but also fires for other uses of \"in\" such as phrases like \"in terms of\", \"in order to\", \"in respect\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-115)\nA/0/251\nThis feature fires when the word \"in\" appears at the beginning of a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-251)\nA/0/353\nThis feature attends to the phrase \"in front of\" and words related to ordering and sequencing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-353)\nA/0/446\nThis feature fires when it sees the preposition \"in\". The feature seems to activate most strongly when \"in\" is used to indicate location, inclusion, or being part of something.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-446)\nA/1/539\n\" in\", tending towards text with a formal tone\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-539)\nA/1/608\n\" In\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-608)\nA/1/1145\n\" in\" (English?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1145)\nA/1/1243\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1243)\nA/1/1833\n\" in\" for medical results\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1833)\nA/1/1956\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1956)\nA/1/2178\n\" In\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2178)\nA/1/2407\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2407)\nA/1/2432\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2432)\nA/1/2467\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2467)\nA/1/2483\nThis feature fires for ``in ____ condition\" or similar phrases describing the state or condition of something. \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2483)\nA/1/2604\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2604)\nA/1/2836\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2836)\nA/1/3058\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3058)\nA/1/3179\nVarious languages: \" in\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3179)\nA/1/3616\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3616)\nA/1/3622\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3622)\nA/1/3721\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3721)\nA/1/3827\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3827)\nA/1/4049\n\" in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4049)\nA/1/47\n\" for\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-47)\nA/1/1224\n'for' in biomedical text\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1224)\nA/1/1419\n\" order\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1419)\nCluster #218\nA/0/473\nThe feature attends to prepositional phrases containing the preposition \"with\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-473)\nA/1/484\n\" from\" in scientific studies\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-484)\nA/1/1046\n\" using\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1046)\nA/1/1679\n\" with\" in science/photography?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1679)\nA/1/2462\n\" by\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2462)\nA/1/4028\n\" via\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4028)\nA/1/3483\n\" with\" in technical contexts? web servers?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3483)\nA/1/2905\n\" with\" in medicine / clinical trials?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2905)\nA/1/1031\n\" by\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1031)\nA/1/645\n\" of\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-645)\nCluster #155\nA/0/335\nThis feature fires when examples start with \"In order to\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-335)\nA/1/627\nWords shortly after sentence-opening \"In\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-627)\nA/1/1451\n\" the\" after \"On\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1451)\nA/1/1511\n\" the\" after \"In\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1511)\nA/1/3315\n\" addition\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3315)\nA/1/2022\n\" from\" in programming/software contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2022)\nA/1/2451\n\" considered\" and broad synonyms in passive voice\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2451)\nA/1/2883\n\" clear\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2883)\nA/1/1245\nWord fragments: passive verbs or some adjectives after \"is\"/word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1245)\nCluster #303\nA/1/381\nAdjectives after \"was\"/\"are\" esp describing people\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-381)\nA/1/1392\nThe feature fires for words that have adjectival suffixes like -ful, -ous, -ible, -able, -al, etc. that indicate judgment or evaluation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1392)\nA/1/2287\n\" crucial\"/\" important\" and synonyms, esp. after \"plays a/an\" (predicting \"role\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2287)\nA/1/3151\n\" unlikely\"/\" impossible\" after \"it is\"/\"it was\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3151)\nA/1/3916\nEmotions, often \" about\" something\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3916)\nA/1/4077\n\" enough\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4077)\nCluster #300\nA/0/193\nThis feature seems to fire the most strongly when the word \"showed\" is present in the text. It also appears to activate slightly for other verbs and phrases related to demonstrating or revealing results, like \"found\", \"observed\", \"evidenced\", and \"obtained\". So this feature likely attends to language describing the results or findings of a study or experiment.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-193)\nA/0/234\nThe feature attends to the verb \"is defined as\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-234)\nA/1/165\n\" shown\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-165)\nA/1/195\n\" thrown\" and similar action verbs likely to receive a direction/preposition\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-195)\nA/1/245\nPast-tense verb about data in prose discussion of code\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-245)\nA/1/920\n\" used\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-920)\nA/1/1371\nAdjectives describing known or established information, esp. \" known\", esp. after \"is\"/\"are\", in medicine\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1371)\nA/1/1414\nPast-tense verb after \"was\" esp. in history/government\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1414)\nA/1/1632\nPast-tense verbs/participles, passive voice, often describing conceptual actions like \"defined\"/\"characterized\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1632)\nA/1/2020\n\" due\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2020)\nA/1/2143\n\" presented\" and similar verbs in science\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2143)\nA/1/2290\n'measured' etc in scientific papers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2290)\nA/1/2567\n\" showed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2567)\nA/1/2578\nThis feature fires when it sees phrases describing a process of manufacturing, depositing, arranging or forming materials, substances or components.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2578)\nA/1/2911\n\" obtained\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2911)\nA/1/3494\nThis feature attends to words used to describe things being required, necessary, or expected.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3494)\nA/1/3513\n\" described\" / \" discussed\" esp in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3513)\nA/1/3523\n\" available\" and some approximate synonyms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3523)\nA/1/3593\nThe feature fires when it encounters words ending in `ed`, `ly`, or `ing`, which are common past-tense, adverb, and progressive verb endings in English.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3593)\nA/1/3737\n\" compared\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3737)\nA/1/3780\n\" associated\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3780)\nA/1/3822\nThe feature fires on passive voice verbs and phrases indicating that something was done.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3822)\nA/1/3081\n\" resulted\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3081)\nCluster #301\nA/1/1796\nPast-tense verbs like \" interested\"/\" participated\", for people in activities, predicting \" in\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1796)\nA/1/2045\n\" arising\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2045)\nA/1/2507\n\" located\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2507)\nA/1/3448\nThis feature fires when it sees a verb form related to the word \"consist.\" It especially activates for forms like \"consisted,\" \"consisting,\" \"consist,\" and \"composed,\" but also somewhat activates for related forms like \"comprised,\" \"comprise,\" \"correspond,\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3448)\nA/1/3708\n\" exist\"/\" occur\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3708)\nA/1/1023\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1023)\nCluster #270\nA/0/12\nThe feature fires when it sees words like \"yet\" and \"only\", suggesting that the feature is paying attention to words highlighting the lack of something or that something hasn’t been completed.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-12)\nA/1/894\nAdverbs that indicate aspects of time\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-894)\nA/1/2329\nAdjectives after \"has\"/\"had\", often size-related metaphorically describing results or plans\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2329)\nA/1/3962\nWords predicting me / myself / I ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3962)\nA/1/2314\n\"Thank\" (maybe religious?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2314)\nA/1/3871\n\" me\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3871)\nA/1/13\n\" thanks\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-13)\nA/1/4060\nPronouns mainly him and them \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4060)\nCluster #164\nA/0/198\nThe feature fires when it sees the word \"something\" and words related to the general concept of something.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-198)\nA/1/458\n\" itself\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-458)\nA/1/1446\n\" here\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1446)\nA/1/1607\n\" nothing\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1607)\nA/1/2021\n\" anyone\"/\" someone\"/\" somebody\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2021)\nA/1/2450\n\" something\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2450)\nA/1/3709\n\" Everything\"/\" everything\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3709)\nA/1/3820\n\"somewhere else\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3820)\nCluster #266\nA/0/237\nThis feature fires on the verb \"do\" and its variations such as \"did\", \"done\", and \"doing.\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-237)\nA/1/302\n\"please\", sometimes capitalized or with leading space\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-302)\nA/1/629\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-629)\nA/1/3009\n\" well\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3009)\nA/1/3206\n\" let\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3206)\nA/1/3317\n\" doing\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3317)\nA/1/3895\n\" do\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3895)\nA/1/4001\n\" need\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4001)\nA/0/492\nThis feature fires when the token \"me\" appears.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-492)\nA/1/1312\nSpiritual, self-help style psychology/charity text, of the sort that would be titled 'Giving is the Key to Happiness'\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1312)\nA/0/102\nThis feature fires when it detects words related to gaining attention or awareness.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-102)\nCluster #279\nA/1/640\nThis feature fires when it sees words about physical interactions such as hands, arms, fingers, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-640)\nA/1/2320\nNouns, esp. abstract, after \"make\" or word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2320)\nA/1/2707\nPronouns after \"make\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2707)\nA/1/3013\nIndirect object pronouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3013)\nA/1/3433\n\" care\"/\" custody\"/etc. after \"take\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3433)\nA/1/3890\n\" together\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3890)\nA/1/417\nObject pronouns after \"tell\" or word forms (predicting \" why\"/\" how\"/etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-417)\nCluster #282\nA/1/1131\n\" able\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1131)\nA/1/1241\nPast participles after \"have\"/\"has\"/\"had\", usually transitive and predicting object pronouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1241)\nA/1/2166\n\" unable\"/\" fail\" esp after \"was\"/\"were\"/etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2166)\nA/1/3224\nThe feature attends to sentences that describe someone's feelings or emotions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3224)\nA/0/330\nThis feature seems to fire on the word \"want\" and its variations such as \"wanted\", \"wanting\", \"wishes\", etc. So the feature is attending to expressions of desire or intent.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-330)\nA/1/1408\n\" want\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1408)\nA/1/2275\n\" trying\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2275)\nCluster #306\nA/0/65\nThis feature is activated when it detects an instance of the phrase \"going to\" or related variations (such as \"gonna\", \"go ahead\", \"go back\", etc.).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-65)\nA/1/1422\n\" bet\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1422)\nA/1/1613\n\" going\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1613)\nA/1/2154\n\" try\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2154)\nA/1/3874\n\" go\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3874)\nA/1/3994\n\" went\" (some \" gone\"/\" goes\"/...\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3994)\nA/1/3\n\"worth\" single-token feature, esp. after \"is\" or word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3)\nA/1/2234\n\" said\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2234)\nCluster #294\nA/0/138\nThe feature seems to fire when expressions of thinking such as \"think\", \"figured\", \"believe\", etc. are used.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-138)\nA/0/480\nThe feature fires for phrases related to knowing or understanding.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-480)\nA/1/585\n\" believe\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-585)\nA/1/775\n\" forget\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-775)\nA/1/1080\nPresent-tense verb after / near \"you\"/\"You\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1080)\nA/1/1234\n\" say\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1234)\nA/1/1334\n\" told\"/\" tells\"/word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1334)\nA/1/1381\n\" know\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1381)\nA/1/1651\n\" sure\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1651)\nA/1/2386\n\" hope\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2386)\nA/1/2420\n\" mean\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2420)\nA/1/2482\n\" means\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2482)\nA/1/2582\n\" show\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2582)\nA/1/2630\n\" see\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2630)\nA/1/2751\n\" heard\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2751)\nA/1/2797\n\" think\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2797)\nA/1/3828\n\" thinking\"/\" thought\" word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3828)\nA/1/3929\nVerbs after \"I\", esp. present-tense\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3929)\nA/0/235\nThe feature fires when it sees dialogue and related tags such as \"he/she said\", \"she replied\", \"he told me\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-235)\nA/1/1119\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1119)\nCluster #322\nA/0/333\nThis feature seems to fire on the base verb \"take\" in a variety of grammatical contexts and tenses.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-333)\nA/0/468\nThis feature fires when it sees verbs related to giving or making (give, makes, made, keeping, etc).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-468)\nA/0/497\nThis feature seems to fire when a form of the verb \"get\" is present. Examples include \"getting\", \"got\", \"gotten\", \"get\", \"gets\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-497)\nA/1/500\n\" keep\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-500)\nA/1/554\n\" take\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-554)\nA/1/1666\n\"make\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1666)\nA/1/1747\n\" taken\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1747)\nA/1/1880\n\" spent\"/\" spend\"/etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1880)\nA/1/2658\nThis feature fires when it sees verb forms related to the word \"bring\". The feature attends to words like bring, brought, bringing, and brings.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2658)\nA/1/3495\n\" put\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3495)\nA/1/3628\n\" give\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3628)\nA/1/3984\n\" lose\" word forms, or rarely \" gain\" word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3984)\nA/1/4025\n\" find\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4025)\nA/1/251\n\" wanted\" and other verbs of volition\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-251)\nCluster #323\nA/0/381\nThis feature fires when it sees verbs and adjectives related to visual perception and appearance such as \"look\", \"seem\", \"stare\", \"glow\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-381)\nA/1/2225\nActive/physical verbs, performed by personal pronouns\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2225)\nA/1/3264\n\" look\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3264)\nA/1/3888\n\" speak\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3888)\nA/1/1075\n\" come\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1075)\nA/1/1943\nThis feature fires on the lemma \"get\" in various forms (gotten, getting, etc).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1943)\nA/1/2933\n\"looked\" and \"looking\" as verbs to describe the act of visually directing one's gaze, as well as more figurative usages indicating scrutiny, seeking. \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2933)\nA/1/1615\n\" killed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1615)\nA/0/58\nThis feature fires when it sees words indicating that someone or something has arrived from elsewhere.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-58)\nA/1/1843\nThe feature activates when it sees words related to beginning, such as began, started, finished, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1843)\nCluster #331\nA/0/440\nThis feature fires when the word \"led\" appears.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-440)\nA/1/648\nThe feature attends to verbs that indicate the role something plays, such as serves, referred, attributed, known, appointed, joined, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-648)\nA/1/809\n\" met\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-809)\nA/1/3816\n\" receive\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3816)\nA/1/4075\n\" followed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4075)\nA/1/4046\n\" directed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4046)\nA/1/3446\n\" led\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3446)\nA/1/3262\nThe feature fires when it sees past tense verbs that describe actions in a formal or official context, such as established, developed, accepted, monitored, advertised, executed, introduced, formed, occupied, manufactured, contemplated, challenged, approved, visited, received, retired, and dragged.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3262)\nA/1/1957\n\" return\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1957)\nA/1/3511\n\" decided\" / \" agreed\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3511)\nA/1/2722\n\" happen\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2722)\nCluster #333\nA/0/60\nThis feature seems to fire on verbs related to physical actions and movements like pull, set, carry, move, cut, burn, freeze, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-60)\nA/1/351\n\" drink\", \" eat\", and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-351)\nA/1/1003\n\" break\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1003)\nA/1/1104\n\" turned\", turns, turning - token\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1104)\nA/1/1156\n\" set\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1156)\nA/1/1838\n\" carried\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1838)\nA/1/2913\n\" pick\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2913)\nA/1/3195\n\" cut\"/\" pull\"/\" knock\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3195)\nA/1/3536\n\" play\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3536)\nA/1/3681\n\" follow\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3681)\nA/1/3713\n\" buy\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3713)\nA/1/2516\n\" talking\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2516)\nA/1/174\n\" read\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-174)\nA/1/215\n\" wait\"/\" waiting\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-215)\nA/1/2374\n\" fall\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2374)\nCluster #332\nA/0/463\nThis feature fires when a sentence contains \"out\" as part of a directional phrasal verb.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-463)\nA/1/86\nPrimarily \" died\", few other past-tense verbs for major life events\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-86)\nA/1/347\n\" grow\"/\" grown\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-347)\nA/1/958\n\" up\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-958)\nA/1/1216\n\" off\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1216)\nA/1/2315\n\" out\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2315)\nA/1/2403\n\" out\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2403)\nA/1/3071\n\" start\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3071)\nA/1/3204\n\" move\" and word forms, of people/things, predicting explanations of how it moved\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3204)\nA/1/3267\n\" back\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3267)\nA/1/3417\n\" away\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3417)\nA/1/3438\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3438)\nA/1/3321\n\" called\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3321)\nA/1/2245\n\" interact\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2245)\nA/1/1027\n\" beginning\"/\" starting\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1027)\nCluster #329\nA/1/1901\n\" living\"/\" staying\" and forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1901)\nA/1/3375\n\" stand\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3375)\nA/1/3612\n\" lie\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3612)\nA/1/1636\nThis feature seems to fire on past tense verbs, especially ones ending in -ed.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1636)\nA/1/1822\nPast-tense verbs for physical actions performed by people\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1822)\nA/1/3139\nWord endings with \"-ing\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3139)\nA/1/552\n\" left\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-552)\nCluster #327\nA/0/32\nThis feature seems to attend to verbs that indicate some sort of change, such as reassign, switch, upgrade, enact, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-32)\nA/1/765\n\" create\"/\" develop\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-765)\nA/1/3370\nAbstract political actions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3370)\nA/1/3945\nThe feature seems to fire on verbs related to concealing or revealing information.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3945)\nA/1/4026\n\" help\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4026)\nA/1/526\n\" agree\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-526)\nA/1/1728\n\" learn\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1728)\nCluster #330\nA/0/263\nThis feature fires when it sees words related to establishing or determining something.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-263)\nA/1/691\n\" provide\" and word forms, usually in an abstract sense\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-691)\nA/1/1721\n\" suffer\" and word forms, sometimes other verbs related to injuries\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1721)\nA/1/2119\n\" seek\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2119)\nA/1/2375\nThis feature attends to words and phrases associated with allowing or granting permission, approval, or access. Examples include \"allow\", \"enable\", \"permit\", \"gives\", \"recommend\", \"consent\", and \"restricting\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2375)\nA/0/496\nThis feature fires when it sees present participle verb forms (-ing) as well as some similar sounding endings like -ington, -ation, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-496)\nA/1/1306\n'choose' token, literal meaning of choice\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1306)\nA/1/2999\n\" love\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2999)\nA/1/2530\n\" ensure\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2530)\nCluster #176\nA/0/2\nThis feature fires on sentences that have the word \"based\" or similar words like \"relying\", \"depend\" and \"insistence\". These sentences often describe something being built or formed from something else.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-2)\nA/1/255\n\" based\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-255)\nA/1/1133\n\" rely\"/\" focus\" word forms etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1133)\nA/1/1983\n\" depending\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1983)\nA/1/2639\nGerunds related to cleaning/physical maintenance\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2639)\nA/1/3953\n\" inspire\"/\" enjoy\"/positive language\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3953)\nA/1/2345\n\" opening\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2345)\nA/1/1024\n\" determine\"/\" ascertain\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1024)\nCluster #287\nA/0/434\nThis feature is activated by examples describing the act of doing work or a job.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-434)\nA/1/617\n\" act\"/\" acts\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-617)\nA/1/785\n\" check\", esp. \" check out\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-785)\nA/1/2746\n\" works\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2746)\nA/1/2787\n\"change\" as a process of transition, development in living organisms, natural systems, social dynamics and human institutions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2787)\nA/1/2825\n\" working\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2825)\nA/1/3491\n\" work\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3491)\nA/1/369\n\" pay\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-369)\nA/1/3793\n\" sat\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3793)\nA/0/56\nThis feature attends to phrases and words related to contacting and getting in touch such as \"Contact us\", \"Call today\", \"Pick up\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-56)\nA/1/3255\n\" Google\"/\"Google\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3255)\nA/0/245\nThe feature fires when it sees case law citations such as \"State v. DefendantName\" and legal language such as \"appellant\" and \"jurisdictional status\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-245)\nA/1/1352\nTokens in the middle or at the end of names in legal contexts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1352)\nA/1/1581\nCapitalized start-of-sentence imperatives/instructions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1581)\nA/1/880\nTitle-case, mainly business/marketing-related words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-880)\nA/1/38\nParty in a court case of \"A vs B\" formatting\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-38)\nCluster #234\nA/1/483\n\" Supreme\" in \"Supreme Court\", often following a U.S. state\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-483)\nA/1/3102\nTitle Case words in names of laws, legal agreements, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3102)\nA/1/3933\n\" Act\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3933)\nCluster #236\nA/1/1601\n\" United\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1601)\nA/1/2494\nCapitalized compass directions esp. \" North\" predicting plausible location names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2494)\nA/1/2719\nDirectional adverb in geography\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2719)\nA/1/3677\n\" News\" and other names of news organizations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3677)\nA/1/686\nTitle Case Nouns esp. undelimited titles in sentences\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-686)\nCluster #237\nA/1/452\n\" Ever\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-452)\nA/1/563\n\" New\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-563)\nA/1/1160\nThis feature fires when it sees black (the color or race) in text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1160)\nA/1/1300\n\" High\" in high school names, e.g. \"Riverdale High School\"; or rarely \" Middle\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1300)\nA/1/1799\n\" Big\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1799)\nA/1/2522\n\" Golden\"/\"Golden\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2522)\nA/1/3476\n\" Good\"/\"Good\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3476)\nCluster #252\nA/0/461\nThis feature fires on proper names containing one or more capital letters, in the context of surrounding text that looks like it could be an article, paper, or other professional writing. Some clues: - The feature consistently has high activation on capitalized words, especially multi-letter abbreviations like NYPD, NETL, CC BY, etc. - It also often fires on locations and organization names like Stonehill College, Aer Lingus, Padova Astronomical Observatory. - The surrounding text has a professional/academic style, mentioning things like a postdoctoral fellowship, the Second World War, International Conferences, etc. So in summary, the feature is detecting proper names and abbreviations, in the context of an academic/professional article or report.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-461)\nA/0/485\nThe feature fires when it sees proper nouns such as names of people, places, organizations, acronyms, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-485)\nA/0/508\nThis feature attends to proper nouns that contain uppercase words or abbreviations in the middle of a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-508)\nA/1/80\nPaper acknowledgements, esp. names of institutions therein\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-80)\nA/1/176\nTitle Case (News Headline)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-176)\nA/1/308\nTitle Case formal model names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-308)\nA/1/419\nThis feature fires on words and phrases related to legal cases, such as \"Case:\", \"Claim No.\", \"Case No.\", \"Court\", \"Order\", \"Form\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-419)\nA/1/701\nCapitalized scientific publications/institutions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-701)\nA/1/1172\nTitle Case words for abstract units of data or relating to finance/payments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1172)\nA/1/1413\nCapitalized terms for organizations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1413)\nA/1/1587\n\" Health\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1587)\nA/1/1703\n\" Law\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1703)\nA/1/2046\nTitle Case Medical terms, e.g. in titles of papers/studies\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2046)\nA/1/2075\n\" War\", esp. in \"World War\", predicting \" II\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2075)\nA/1/2126\n\" Down\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2126)\nA/1/2232\nCapitalized, esp. last, words in formal names of organizations/awards/etc. but before the \"kind of thing\" it is (e.g. not \" Committee\" or \" Association\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2232)\nA/1/2300\n\" Time\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2300)\nA/1/2772\nTitle-case words, esp. \" Dream\", esp. used in a \"compound word\" name like \"Dreamcast\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2772)\nA/1/3166\n\" National\" in Title Case organization names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3166)\nA/1/3307\nAcademic Institution Name\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3307)\nA/1/4066\n\" World\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4066)\nCluster #251\nA/0/99\nThis feature fires when proper nouns are present, especially locations like cities and states.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-99)\nA/1/293\nNames of roads and streets\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-293)\nA/1/306\nCountries or continents, capitalized\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-306)\nA/1/687\nPlaces with bipartite names (_ Creek, _ Falls)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-687)\nA/1/2566\n\" City\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2566)\nA/1/3447\nMajor American city names, predicting locations/things in them, e.g. \" City\" or \" County\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3447)\nA/1/4018\n\" York\" in \"New York\", predicting things in New York, e.g. \" City\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4018)\nA/1/1985\nMusic/Dance/Broadway/Movies named entities\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1985)\nA/1/1752\nEnds of Title Case company/brand names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1752)\nA/1/3869\n\" Part\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3869)\nA/1/2904\n\" General\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2904)\nCluster #229\nA/0/337\nThis feature fires on italics or book/article titles.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-337)\nA/1/1665\nTitle Case words / word ends in book titles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1665)\nA/1/2851\nUnderscore-delimited names/words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2851)\nA/1/3861\nCapitalized royal/religious titles (\" Lord\", \" \"King\", etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3861)\nA/1/2070\nMostly ends of proper nouns / names of people, predicting past-tense verbs they might perform\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2070)\nA/0/297\nThis feature attends to names ending in a, ie, or y.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-297)\nA/1/3456\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3456)\nA/1/3936\n\" Lock\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3936)\nA/1/2705\n\" Out\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2705)\nA/1/2460\nTitle Case location names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2460)\nA/1/2312\nThe feature fires when words contain \"wood\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2312)\nA/1/346\nCommon first names, predicting past-tense verbs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-346)\nA/1/394\n\" Inc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-394)\nCluster #357\nA/0/445\nThis feature fires on English names in text, particularly last names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-445)\nA/1/123\nNames of authors/locations/publishers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-123)\nA/1/646\nThis feature fires when it sees words containing the substring \"ley\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-646)\nA/1/672\nCommon English last names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-672)\nA/1/1084\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1084)\nA/1/1190\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1190)\nA/1/2293\nThe feature seems to fire on Spanish surnames ending in vowels such as a, e, i, o.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2293)\nA/1/2703\nEnds of names, particularly people, in English\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2703)\nA/1/2918\nRoman/Greek proper names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2918)\nA/1/2973\nNames of people or final tokens thereof\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2973)\nA/1/3098\nLast token of surname, in legal descriptions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3098)\nA/1/3240\n\" Trump\" (Donald Trump) / some other U.S. presidents\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3240)\nA/1/3664\nEnglish, esp. British, place names, predicting \"shire\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3664)\nA/1/3818\nTokens in the middle or at the end of a name that looks like a location, predicting e.g. \" County\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3818)\nA/1/3060\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3060)\nCluster #354\nA/1/17\n\"er\" and \"ter\", usually near ends of words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-17)\nA/1/101\n\"er\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-101)\nA/1/525\nEnding of a person's name\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-525)\nA/1/1240\nThis feature fires for words whose final syllable sounds like \"ore\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1240)\nA/1/2041\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2041)\nA/1/2123\nThe feature attends to tokens that end or have a \"y\" sound.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2123)\nCluster #364\nA/0/370\nThis feature attends to male first names ending with \"n\" (and variations) such as Kevin, Steven, Ken, John, Glenn, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-370)\nA/1/124\nTypical English honorifics \"Mr\"/\"Ms\"/\"Mrs\"/\"Miss\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-124)\nA/1/1244\nCommon English first names, esp. male, predicting last names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1244)\nA/1/2697\n\" Dr\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2697)\nA/1/891\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-891)\nCluster #363\nA/0/314\nThe feature fires when there are names or words ending with \"in\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-314)\nA/1/1592\nThis feature attends to foreign names and words, especially German ones. \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1592)\nA/1/2036\nParts of names, esp. endings and esp. people's last names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2036)\nA/1/3717\nThis feature fires on words that end in \"ar\" such as \"Dakar\", \"Hoyer\", and on words that end in \"al\" such as \"Scalial\". It also seems to activate slightly on some words ending in \"er\" such as \"Birch\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3717)\nCluster #365\nA/1/1456\n\" Tim\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1456)\nA/1/3250\n\" Christ\"/\"Christ\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3250)\nA/1/3970\n\" Ben\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3970)\nA/1/1969\nTokens in middle or end of names, esp. location names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1969)\nA/1/1805\nEnds of names or usernames, and \"gate\" in \"enronXgate\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1805)\nA/1/2712\n\"e\" in the end of last names of people\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2712)\nA/1/1056\nConjunctions/prepositions near Indian locations/government\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1056)\nA/1/231\nTransliterated Arabic names of people?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-231)\nA/1/1093\nBeginnings of names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1093)\nCluster #362\nA/0/457\nThis feature fires on words and names ending in \"a\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-457)\nA/1/1053\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1053)\nA/1/1827\nThis feature appears to activate on syllables and sounds at the end of words like \"ah\", \"uh\", and \"ih\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1827)\nA/1/1926\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1926)\nA/1/2136\nThis feature attends to Japanese, Korean, Tamil and Hindi names and words.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2136)\nA/1/3256\nMid- or end-word \"as\" or \"is\" fragments\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3256)\nA/1/3501\nMostly \"la\" but various non-English word/name endings ending in \"a\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3501)\nCluster #361\nA/0/260\nThis feature attends to proper nouns ending in \"el\", \"en\" or \"er\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-260)\nA/1/839\n\"in\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-839)\nA/1/939\nParts of a large variety of various names (Korean, Japanese, Persian, Yoruba...)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-939)\nA/1/1305\nSingle-lowercase-letter tokens, e.g. \"b\"/\"l\"/\"n\"/\"m\"/\"k\", inside capitalized words and names with unusual spellings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1305)\nA/1/3842\nThis feature fires when it sees names or words ending in -en.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3842)\nA/1/1502\nPeople's names, particularly Spanish/Latin-American\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1502)\nA/1/1848\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1848)\nCluster #390\nA/0/481\nThis feature attends to the letters \"ul\" surrounded by consonants.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-481)\nA/1/280\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-280)\nA/1/411\nTokens before German name suffixes like -berg, -stein, -reich, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-411)\nA/1/856\n~2-letter tokens after the start of a title-cased name, esp. \"ul\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-856)\nA/1/865\nToken after \" Ca\", either name continuation or number/punctuation for Calcium\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-865)\nA/1/1354\n\"ac\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1354)\nA/1/1417\nTokens in a capitalized word esp. after \"Con\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1417)\nA/1/1454\nTokens in a capitalized word esp. after \"Sch\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1454)\nA/1/1687\n\" McG\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1687)\nA/1/1962\nTokens in names, esp. \" Lind\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1962)\nA/1/2196\nThis feature fires when it sees words containing double consonant letters and diphthongs in Spanish.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2196)\nA/1/2418\nSpace + capital letter after a first name, either a middle initial or starting a last name\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2418)\nA/1/3283\nThe feature appears to fire on proper nouns, especially names, that start with hard \"k\" or \"c\" sounds.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3283)\nA/1/3488\nMid-title-case-name tokens esp. \"it\"/\"ac\" after \" Cap\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3488)\nA/1/3926\n\" Jr\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3926)\nA/1/1602\n\"ist\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1602)\nA/1/904\n\"mb\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-904)\nCluster #35\nA/0/189\nThis feature fires for two or three initials followed by a period, which is a common way of citing author's names in academic papers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-189)\nA/1/509\n\" Rev\" in in citation-like things, most prominently \"Phys. Rev.\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-509)\nA/1/1291\nAdvanced math book titles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1291)\nA/1/2369\nFirst initial of author in citations (usually S)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2369)\nA/1/3075\nCapital initials and initial capital letters of names, esp. in bibliographies, e.g. \"S\" in \"S. S. Smith\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3075)\nA/1/3839\nMiddles and ends of European last names, esp. in bibliographies\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3839)\nA/1/3694\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3694)\nCluster #353\nA/1/1153\nGerman: ch?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1153)\nA/1/2002\n\"av\"/\"ov\" and other mid-word/name tokens with \"v\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2002)\nA/1/2947\n\"ik\" in last/first names from Slavic or Eastern European origins\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2947)\nA/1/3954\nThe feature fires when it sees words with the \"ang\" phoneme in them, such as bang, rang, fang, anguish, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3954)\nA/1/965\nSanskrit, Arabic, and Indic: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-965)\nA/1/3960\n\" Phys\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3960)\nCluster #351\nA/0/230\nThe feature seems to attend to foreign names and terms, primarily names from South Asian languages such as Sanskrit.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-230)\nA/1/1028\nJapanese: traditional names written in Romaji\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1028)\nA/1/1448\nPinyin-transliterated Chinese names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1448)\nA/1/1579\nRussian: Names connected to the country, not necessarily ethnically Russian\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1579)\nA/1/3639\nEnglish transliterations of Indian names?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3639)\nA/1/683\nWord component \"ost\" or \"st\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-683)\nA/1/3619\nThis feature attends to the suffix -ill and names ending in -ill.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3619)\nA/1/1447\nSpace + single capital letter tokens, typically in title case\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1447)\nA/1/1919\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1919)\nA/1/2291\n\"am\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2291)\nA/1/2621\nThis feature fires on words that end in \"ot\". This includes words like \"faggot\", \"campot\", \"expert\", \"caput\", etc. It seems to especially activate on these endings when they appear at the end of a line, like in lists or at the end of a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2621)\nA/1/384\nSpanish first names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-384)\nA/1/2786\n\" Gi\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2786)\nA/1/4048\nEuropean languages: \"az\"/\"iaz\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4048)\nA/1/778\n\" Man\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-778)\nCluster #391\nA/0/345\nThis feature activates for words containing the sound \"cap\" (cap, capacitance, capacitors, etc).\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-345)\nA/1/478\n\" Mon\"/\"Mon\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-478)\nA/1/1296\n\" Bon\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1296)\nCluster #389\nA/1/1836\n\" Santa\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1836)\nA/1/2304\n\" Saint\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2304)\nA/1/2660\n\" San\"/\"San\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2660)\nA/0/383\nThis feature appears to fire on names and places, particularly with Spanish origin such as Benigno Aquino, Cristo Yacente, María de los Milagros, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-383)\nA/1/2900\n\" St\"/\"St\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2900)\nA/1/1732\n\"enn\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1732)\nA/1/3265\n\"und\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3265)\nA/0/236\nThis feature fires for the prefix abbreviation 'St.', most commonly as part of location names such as cities and states.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-236)\nA/1/3638\nThis feature fires when it sees names with Polish spelling, including endings like -ski, -owicz, etc. It also activates for some normal English words that end in -ow like \"shadow\", but definitely has a bias for Polish names.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3638)\nA/1/1886\n\" Math\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1886)\nA/1/210\n\" Ann\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-210)\nA/1/1803\nPinyin-transliterated Chinese names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1803)\nA/1/2260\nThis feature attends to words that end in -ur.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2260)\nA/1/1329\n\" Bow\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1329)\nA/1/1142\nThis feature fires on words that have the \"-ol\" sound in them. Some examples include \"Rolfe\", \"Roller\", \"Rolland\", \"Rolipram\", \"Rolens\", etc. While it sometimes also picks up on other words that end in \"-ol\" like \"control\" or \"alcohol\", there is a clear bias for names and words where the \"-ol\" sound is more emphasized. The feature does not seem to care about the semantics or context, just attending to any word with that phonetic ending.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1142)\nA/1/2818\n\" Van\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2818)\nA/1/3108\n\" Cast\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3108)\nA/1/2479\n\" Fran\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2479)\nA/1/4047\n\"ar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4047)\nA/1/4038\n\" Com\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4038)\nA/1/1533\nMid-word \"og\" or other fragments ending with \"g\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1533)\nA/1/909\n\" Am\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-909)\nA/1/2160\n\" Ok\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2160)\nA/1/2538\n\" Met\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2538)\nA/1/1212\n\" Jo\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1212)\nCluster #392\nA/0/223\nThis feature fires when a word includes the syllable \"rog\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-223)\nA/1/58\n\" Hen\", handful of \" Ken\"/\" Len\"/name starts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-58)\nA/1/431\n\" Med\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-431)\nA/1/594\n\" Ang\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-594)\nA/1/696\n\" Ser\"/\"Ser\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-696)\nA/1/790\n\" Mac\"/\"Mac\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-790)\nA/1/836\n\" Mar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-836)\nA/1/924\n\"Mont\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-924)\nA/1/1180\n\" Mur\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1180)\nA/1/1384\n\" Mad\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1384)\nA/1/1519\nBeginnings of proper nouns esp. \" Em\"/\" Ham\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1519)\nA/1/1853\n\" Er\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1853)\nA/1/1890\nThe feature attends to words with /məl/, /i/, and /ɑː/ sounds (such as Mal-, I-, A-) in the middle or end of words.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1890)\nA/1/1948\n\" Brit\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1948)\nA/1/1984\n\" Mc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1984)\nA/1/2157\n\" Mer\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2157)\nA/1/2527\n\" Mor\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2527)\nA/1/2627\n\" Har\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2627)\nA/1/2644\n\" Luc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2644)\nA/1/2809\n\" Rem\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2809)\nA/1/2981\n\" Bur\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2981)\nA/1/3185\n\" Jan\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3185)\nA/1/3424\n\" Mah\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3424)\nA/1/3527\n\" Mag\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3527)\nA/1/3562\n\" Pet\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3562)\nA/1/3673\n\" Ed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3673)\nA/1/3702\n\" Cor\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3702)\nA/1/3794\n\" Pat\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3794)\nA/1/3802\n\" Mart\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3802)\nA/1/2003\n\" Az\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2003)\nA/1/2261\n\" Bel\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2261)\nA/1/1258\n\" Hor\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1258)\nA/1/3452\n\" Sim\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3452)\nA/1/2052\n\" Sign\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2052)\nA/1/2732\n\"it\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2732)\nCluster #393\nA/0/382\nThe feature attends to words ending in \"-al\", especially when it is part of a proper noun.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-382)\nA/0/471\nThis feature fires on words that contain the syllable \"terra\" such as territory, terracing, Terran, and terraced.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-471)\nA/1/147\n\" sal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-147)\nA/1/153\n\" Wil\" in names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-153)\nA/1/352\n\" Car\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-352)\nA/1/548\n\" Col\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-548)\nA/1/566\n\"ou\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-566)\nA/1/984\n\" Ter\"/\"Ter\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-984)\nA/1/1219\n\" Sac\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1219)\nA/1/1536\n\"mar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1536)\nA/1/1621\n\" ger\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1621)\nA/1/1630\n\" Sand\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1630)\nA/1/1730\n\" Par\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1730)\nA/1/1733\n\" Sy\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1733)\nA/1/1896\n\" El\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1896)\nA/1/1925\n\" Ban\"/\"Dun\"/\" Dun\", esp. beginning names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1925)\nA/1/1952\n\" Kar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1952)\nA/1/2282\n\" Bal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2282)\nA/1/2340\n\" ar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2340)\nA/1/2352\n\" Anal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2352)\nA/1/2517\n\" Gal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2517)\nA/1/2573\n\" Hel\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2573)\nA/1/2704\n\"ij\" in \"ijerph\" citations (\"@B1-ijerph-...\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2704)\nA/1/2845\n\" Pol\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2845)\nA/1/2875\n\" Al\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2875)\nA/1/3047\n\" Cal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3047)\nA/1/3062\n\" Grand\"/\"Grand\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3062)\nA/1/3246\n\" Ber\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3246)\nA/1/3331\n\" Far\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3331)\nA/1/3404\n\" Vol\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3404)\nA/1/3521\n\" Ly\" and similar usually capitalized word beginnings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3521)\nA/1/3522\n\" Ar\" and similar usually capitalized word beginnings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3522)\nA/1/3800\n\" Sher\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3800)\nA/1/3876\n\" Del\"/\"Del\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3876)\nA/1/3999\n\" Wal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3999)\nA/1/4065\n\" Val\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4065)\nA/1/2150\n\"leg\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2150)\nA/1/2090\n\" bar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2090)\nA/1/942\nMid-capitalized-word tokens ending in \"t\", sometimes \"ct\", predicting \"orial\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-942)\nA/1/1532\n\" Port\"/\"Port\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1532)\nA/1/2515\nVer\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2515)\nA/1/3274\n\" Mat\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3274)\nCluster #382\nA/1/2940\n\" Dam\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2940)\nA/1/2953\n\" Mot\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2953)\nA/1/3126\n\" Sav\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3126)\nA/1/1439\n\" Gen\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1439)\nCluster #381\nA/1/1037\n\" Ant\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1037)\nA/1/1361\n\" Ind\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1361)\nA/1/2008\n\" ed\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2008)\nA/1/2547\nMostly \" Ent\", some \"Jac\"/\"Vent\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2547)\nA/1/3295\n\" Vis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3295)\nA/1/502\n\"os\" ending a word (often Spanish word or name)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-502)\nCluster #371\nA/1/595\n\" Res\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-595)\nA/1/888\n\" Den\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-888)\nA/1/1483\n\" Nat\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1483)\nA/1/2197\n\" pul\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2197)\nA/1/3725\n\" Mus\"/\"Mus\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3725)\nA/1/3427\nWord fragments predicting \"let\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3427)\nA/1/737\n\" Pan\"/\" pan\"/\"Pan\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-737)\nA/1/1960\n\" Pal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1960)\nA/1/3396\n\" Fre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3396)\nA/1/1508\n\" Sil\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1508)\nA/1/763\n\" Appro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-763)\nA/1/276\n\" sol\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-276)\nA/1/3200\n\" Di\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3200)\nA/1/2142\n\" Re\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2142)\nA/1/1550\n\"co\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1550)\nA/1/3379\n\" La\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3379)\nCluster #384\nA/0/365\nThis feature attends to names or words starting with 'Pe' or 'Se'.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-365)\nA/1/203\n\" Pe\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-203)\nA/1/274\n\" Co\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-274)\nA/1/425\n\" Me\"/\"Me\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-425)\nA/1/893\n\" Be\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-893)\nA/1/1401\n\" Bre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1401)\nA/1/1487\n\" Se\"/\"Se\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1487)\nA/1/1566\n\" Le\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1566)\nA/1/1794\n\" Ke\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1794)\nA/1/2039\n\"ste\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2039)\nA/1/2159\n\" Che\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2159)\nA/1/3109\n\" Ste\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3109)\nA/1/3734\n\" Ne\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3734)\nA/1/2508\nThis feature attends to words ending in \"o\" or containing \"cho\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2508)\nA/1/117\n\" De\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-117)\nCluster #385\nA/0/176\nThis feature attends to tokens that start with Sa or Ba.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-176)\nA/1/722\nThe feature attends to words that start with \"Ro\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-722)\nA/1/1591\n\" Ha\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1591)\nA/1/1628\n\" Bo\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1628)\nA/1/2258\n\" Ca\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2258)\nA/1/2343\n\" Sa\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2343)\nA/1/2426\n\"ri\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2426)\nA/1/3083\n\" po\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3083)\nA/1/3458\n\" ca\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3458)\nA/1/3658\n\" Lo\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3658)\nA/1/590\n\" Ki\" and other space, capitalized-consonant, vowel tokens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-590)\nCluster #383\nA/0/361\nThe feature seems to fire when words start with \"tra\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-361)\nA/1/200\n\" Fra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-200)\nA/1/423\n\"ra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-423)\nA/1/667\n\" Tre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-667)\nA/1/1276\n\" fe\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1276)\nA/1/1643\n\" Bro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1643)\nA/1/1685\n\"Tro\"/\" Tro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1685)\nA/1/1855\n\" pri\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1855)\nA/1/2977\n\" Cro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2977)\nA/1/3667\n\" Bra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3667)\nA/1/3705\n\" Tra\"/\" tra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3705)\nA/1/3798\n\" Dra\"/\" dra\"/\"Dra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3798)\nA/1/3803\n\"ro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3803)\nA/1/3819\n\"du\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3819)\nA/1/3832\n\" Gra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3832)\nA/0/62\nThis feature seems to fire on co- words, pairs that share a prefix such as co-author, co-founder, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-62)\nA/1/1387\n\" co\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1387)\nA/1/2598\n\" tri\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2598)\nA/1/3443\n\" Ju\"/\"Ju\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3443)\nA/1/1248\n\" fre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1248)\nA/1/4003\n\" ri\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4003)\nA/1/2968\n\" cal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2968)\nA/1/4093\nMid-word \"iqu\", start-of-word \"nu\", and other \"-u\" tokens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4093)\nA/0/506\nThe feature attends to occurrences of \"et al\" in citations or lists of authors.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-506)\nA/1/2027\n\" et\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2027)\nCluster #335\nA/1/636\nStart-of-word \"vari\"/\"propri\", predicting \"ety\" and forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-636)\nA/1/744\n\" fore\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-744)\nA/1/3690\n\" dise\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3690)\nA/1/3489\n\" al\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3489)\nA/0/131\nThis feature fires when it sees references to scientific papers or documents. It looks for things like author last names followed by \"et al.\", citations like [CR17], brackets for citations like (@CR17), numbered references in the text, and other stylistic patterns common in academic writing.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-131)\nA/1/2723\n\" hal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2723)\nA/1/1333\n\" Cy\"/\"Cy\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1333)\nA/1/439\n\" Ad\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-439)\nA/1/3683\n\" per\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3683)\nA/1/304\n\"Per\", as a word prefix\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-304)\nA/1/2085\n\" en\", generally not Spanish\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2085)\nA/1/2863\n\" Sur\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2863)\nA/1/285\n\" En\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-285)\nA/1/132\n\" Ge\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-132)\nA/1/270\nFrench or German: sur\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-270)\nA/1/412\n\" ap\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-412)\nA/1/2082\n\" Pro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2082)\nA/1/53\n\" Impro\"/\" impro\" with some \" Retrie\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-53)\nA/1/3154\n\" hy\"/\"Hy\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3154)\nCluster #315\nA/0/369\nThis feature fires when it sees the Latin prefix \"con-\" or \"com-\" meaning \"together, with\" at the beginning of English words. Some less frequent languages like French and Italian are also activated when they contain words starting with \"con-\" or \"com-\". The feature does not activate on other standalone letters like \"c\" or \"o\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-369)\nA/1/54\n\"con\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-54)\nA/1/261\n\" Con\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-261)\nA/1/689\n\" ex\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-689)\nA/1/1460\n\" con\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1460)\nA/1/2975\n\" Ex\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2975)\nA/1/3111\n\" trans\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3111)\nCluster #324\nA/0/54\nThis feature attends to prefixes dis- and in-, which are often used to negate words or express the opposite meaning.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-54)\nA/0/332\nThis feature attends to words starting with \"re\" that are often prefixes or suffixes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-332)\nA/0/410\nThis feature seems to attend to words with the \"un-\" prefix, including misspellings or errors such as \"untrained\" or \"uncertificated\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-410)\nA/1/998\n\" Un\"/\"Un\", esp. right after EOT or in news/politics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-998)\nA/1/2047\n\" mis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2047)\nA/1/2646\n\" re\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2646)\nA/1/3067\n\" pro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3067)\nA/1/3451\n\" Dis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3451)\nA/1/3998\n\" un\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3998)\nA/1/1322\n\"Pre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1322)\nA/1/2149\n\"Tran\"/\" Tran\", predicting \"script\" and forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2149)\nA/1/1433\n\" dis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1433)\nA/1/1368\n\" Des\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1368)\nCluster #325\nA/0/422\nThis feature fires when it sees prefixes like: pre-, non-, intra-, inter-.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-422)\nA/1/1055\n\" inter\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1055)\nA/1/1171\n\" sub\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1171)\nA/1/1515\n\" pre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1515)\nA/1/1660\n\" micro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1660)\nA/1/2203\n\" bio\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2203)\nA/1/2412\n\" anti\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2412)\nA/1/2445\n\" hyper\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2445)\nA/1/2633\n\" non\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2633)\nA/1/2677\n\" intra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2677)\nA/1/3389\n\" super\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3389)\nA/1/3855\n\" Sub\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3855)\nA/1/4020\n\" multi\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4020)\nA/1/3479\n\" mid\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3479)\nCluster #372\nA/0/50\nThe feature seems to fire when it sees words ending with \"res\". This includes common English words like \"resist\", \"rescue\", \"resign\", as well as words from other languages like French (\"Desrosiers\"), German (\"Fes\"), Spanish (\"Residuos\") etc. The feature is picking up on a specific suffix/ending shared across languages rather than attending to any deeper semantic or stylistic features.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-50)\nA/1/1092\n\" Es\"/\" es\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1092)\nA/1/3344\n\" Ass\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3344)\nA/1/3545\nStart-of-word \"Sus\"/\" Sus\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3545)\nA/1/3620\n\" tele\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3620)\nA/1/2461\n\" neuro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2461)\nA/1/443\n\" electro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-443)\nA/1/464\n\" Gu\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-464)\nA/1/294\n\" mal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-294)\nA/1/746\n\" Nut\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-746)\nCluster #374\nA/1/551\nMid-word token like \"ib\", containing vowel followed by \"b\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-551)\nA/1/1690\n\" Lib\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1690)\nA/1/1719\n\" Ab\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1719)\nA/1/3892\n\" cor\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3892)\nA/1/3769\n\" McK\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3769)\nA/1/3570\n\" res\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3570)\nA/1/1196\n\" Uns\"/\"Uns\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1196)\nA/1/3745\n\" Fer\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3745)\nA/1/2425\nMostly \" Benef\"/\" Jud\" predicting \"iciary\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2425)\nA/1/2959\n\" ammon\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2959)\nA/1/989\n\" ep\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-989)\nA/1/3915\n\" ant\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3915)\nA/1/1372\n\"colog\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1372)\nA/1/2829\n\" Art\"/\" art\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2829)\nA/1/1588\n\" min\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1588)\nA/1/2649\n\"ob\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2649)\nA/0/338\nThis feature attends to the string \"ab\" occurring anywhere in a word, regardless of case.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-338)\nA/1/1942\n\" ab\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1942)\nA/1/3260\n\" ob\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3260)\nA/1/3599\n\" imm\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3599)\nA/1/1481\n\"ap\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1481)\nA/1/3492\n\" Emp\"/\"Emp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3492)\nCluster #380\nA/0/478\nThis feature attends to words related to admissions such as \"admission\", \"admissions\", \"admissible\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-478)\nA/1/1605\n\" Bot\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1605)\nA/1/1620\n\" Cap\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1620)\nA/1/1688\n\" rect\" and \"scar\", predicting \"ify\" and forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1688)\nA/1/1708\n\" mod\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1708)\nA/1/1907\n\" gall\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1907)\nA/1/2194\n\" Log\"/\" log\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2194)\nA/1/2249\n\" Net\"/\" net\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2249)\nA/1/2803\n\" Reg\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2803)\nA/1/3087\n\" col\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3087)\nA/1/3132\nThis feature seems to fire the most on words that have the -ion or -ive suffixes and sometimes -ment. Specifically, it looks for words like communicative, classification, adaptable, detachment, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3132)\nA/1/3198\n\" ad\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3198)\nA/1/3465\n\" Bet\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3465)\nA/1/2763\n\" cur\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2763)\nA/1/3496\n\" reg\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3496)\nA/1/344\n\" amp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-344)\nA/1/761\n\" App\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-761)\nCluster #396\nA/0/279\nThe feature seems to fire when it sees words related to computing and computers, such as \"computed\", \"computationally\", \"computer\", etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-279)\nA/1/115\n\" ag\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-115)\nA/1/221\n\" inf\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-221)\nA/1/258\n\" ext\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-258)\nA/1/750\n\" rep\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-750)\nA/1/771\n\" amb\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-771)\nA/1/1539\n\" imp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1539)\nA/1/1650\n\" ref\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1650)\nA/1/1835\n\" emb\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1835)\nA/1/2081\n\" exp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2081)\nA/1/2700\n\" Comp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2700)\nA/1/3137\n\" conf\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3137)\nA/1/3225\n\" Dep\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3225)\nA/1/3279\n\" Sup\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3279)\nA/1/3328\n\" Imp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3328)\nA/1/3850\n\" Ref\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3850)\nA/1/1202\n\" Ev\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1202)\nA/1/1673\n\" cent\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1673)\nA/0/77\nThis feature seems to identify words that are commonly used as suffixes in English. Some examples of suffixes it highlights include: -tion, -ment, -phy, -ing, -ed, -er, -ive, -ous, -ful, -ly, -ible, -able, -ure, -ship, -hood, -ness, -ity, -ence, -ant, -ial, -ary, -ery, -ory, -ion, -ity, -ward, -wise, -less, -ful, -ic, -ist, -ive, -ous. So to summarize, this feature seems to fire on English suffixes.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-77)\nA/1/2596\n\"Rep\", esp. right after EOT and short for (U.S.) Representative\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2596)\nCluster #388\nA/0/423\nThe feature seems to fire for tokens containing the word \"hop\" (and related words like \"hip\"), specifically in the context of hip-hop music.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-423)\nA/1/319\n\" bur\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-319)\nA/1/599\n\" wrink\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-599)\nA/1/601\n\" scal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-601)\nA/1/1091\n\" hel\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1091)\nA/1/1360\nMostly \" pot\" with some \" bull\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1360)\nA/1/1913\n\" bount\"/\" mas\"/\" bal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1913)\nA/1/3356\n\" bug\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3356)\nA/1/3512\n\" pop\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3512)\nCluster #397\nA/1/1554\n\" Int\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1554)\nA/1/1571\n\" def\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1571)\nA/1/2339\n\" Def\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2339)\nCluster #386\nA/1/208\n\" compress\", \"uggest\", and similar penultimate word tokens, predicting suffixes like \"ive\"/\"or\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-208)\nA/1/496\n\" cult\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-496)\nA/1/2458\n\" apost\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2458)\nA/1/2828\nMostly \" gam\" (\"gamers\"/\"gamblers\"/\"gamification\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2828)\nA/1/883\n\" Cont\"/\"Cont\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-883)\nA/1/1906\n\" cat\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1906)\nCluster #398\nA/1/808\n\" Ret\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-808)\nA/1/3043\n\" downt\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3043)\nA/1/3974\n\" unt\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3974)\nA/1/3637\n\" post\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3637)\nA/1/3144\nStart-of-word \"elect\", sometimes capitalized and/or with leading space\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3144)\nCluster #288\nA/1/671\n\" rest\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-671)\nA/1/949\n\" press\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-949)\nA/1/1138\n\" cross\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1138)\nA/1/1428\n\" camp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1428)\nA/1/2114\n\" trip\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2114)\nA/1/1064\n\" ent\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1064)\nA/1/1159\n\" disc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1159)\nA/1/1872\n\" refract\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1872)\nCluster #395\nA/0/9\nThis feature seems to fire for words that start with 'sc' and 'acc' like scenery, scuff, scuttle, accents, accelerate, accidents, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-9)\nA/1/6\n\" rec\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-6)\nA/1/534\n\" enc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-534)\nA/1/538\n\" inc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-538)\nA/1/882\n\"oc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-882)\nA/1/1347\n\" Ac\"/\" ac\", sometimes in health/science\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1347)\nA/1/1625\n\" Occ\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1625)\nA/1/1977\n\" dec\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1977)\nA/1/2055\n\" prof\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2055)\nA/1/2795\n\" dev\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2795)\nA/1/4067\n\" prov\"/\" inv\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4067)\nA/1/520\n\"sk\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-520)\nCluster #316\nA/1/2872\n\" inst\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2872)\nA/1/3290\n\" Sc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3290)\nA/1/4070\n\" Ph\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4070)\nA/1/3846\n\" Ins\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3846)\nCluster #309\nA/0/194\nThis feature fires for words ending with \"st\" such as past, post, pist, wast, etc..\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-194)\nA/0/399\nThe feature fires when it sees words beginning with pl, bl, fl or tr.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-399)\nA/0/452\nThis feature seems to fire for examples of words or names that start with \"Sh\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-452)\nA/1/36\n\" Sk\" and few other Title Case word starts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-36)\nA/1/197\n\" cl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-197)\nA/1/205\n\" fl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-205)\nA/1/435\n\" Sh\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-435)\nA/1/550\n\" Str\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-550)\nA/1/660\n\" ch\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-660)\nA/1/688\n\" Fl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-688)\nA/1/708\n\" Cr\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-708)\nA/1/844\n\" pl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-844)\nA/1/859\n\" Wh\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-859)\nA/1/889\n\" sk\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-889)\nA/1/1014\n\" sw\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1014)\nA/1/1185\n\" Pl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1185)\nA/1/1187\n\" wh\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1187)\nA/1/1199\n\" cr\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1199)\nA/1/1398\n\" Th\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1398)\nA/1/1694\n\" sn\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1694)\nA/1/1974\n\" Sm\"/\"Sm\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1974)\nA/1/2017\n\" sl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2017)\nA/1/2091\n\" sh\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2091)\nA/1/2442\n\" Ch\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2442)\nA/1/2503\n\" Sw\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2503)\nA/1/2579\n\" th\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2579)\nA/1/2613\n\" bl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2613)\nA/1/2638\n\" Fr\"/\" fr\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2638)\nA/1/2898\n\" str\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2898)\nA/1/2902\n\" gr\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2902)\nA/1/3191\n\" st\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3191)\nA/1/3327\n\" sp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3327)\nA/1/3346\n\" Sch\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3346)\nA/1/3546\n\" Sp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3546)\nA/1/3598\n\" Br\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3598)\nA/1/3806\n\" Cl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3806)\nA/1/3860\n\" tr\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3860)\nA/1/3886\n\" Sl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3886)\nA/1/4056\n\" gl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4056)\nA/1/3208\n\" sc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3208)\nA/1/2939\n\" ev\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2939)\nA/1/1239\nThe feature appears to fire for words containing the phonetic \"arc\" such as \"arcades\", \"autoradiography\", \"choreocutaneous\", \"echocardiography\", \"hydrocephalus\", \"microautoradiography\", \"photocoagulation\" and \"ultrasonocardiography\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1239)\nCluster #387\nA/1/681\n\" vag\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-681)\nA/1/1825\n\" fib\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1825)\nA/1/3363\n\" mac\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3363)\nA/1/3500\n\" vis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3500)\nA/1/1614\n\" pent\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1614)\nA/1/516\n\"act\" as part of a word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-516)\nA/1/3414\n\" hall\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3414)\nA/1/3736\nMostly \" meth\", some \" phen\" and other chemical prefixes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3736)\nCluster #159\nA/1/219\nThis feature fires for words containing the letter combination \"it\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-219)\nA/1/1737\n\" gly\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1737)\nA/1/2997\n\" glut\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2997)\nA/1/3432\n\" coll\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3432)\nA/1/1920\n\" ax\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1920)\nCluster #379\nA/0/91\nThe feature seems to attends to words containing \"qu\", particularly at the beginning of the word. This includes English words like \"quite\", \"quote\", and \"quibble\", as well as non-English words like \"Québec\" and \"quels\". The feature also fires for some words containing \"qu\" in the middle, like \"accommodating\", but more weakly.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-91)\nA/1/196\n\" qu\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-196)\nA/1/826\n\" gu\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-826)\nA/1/1478\n\"Qu\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1478)\nA/1/1970\n\" Vo\"/\" vo\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1970)\nA/1/3169\n\" equ\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3169)\nA/1/3670\n\" Mult\"/\" mult\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3670)\nA/1/4016\n\"qu\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4016)\nA/1/1485\n\" poll\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1485)\nA/1/3979\n\" ir\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3979)\nA/1/3249\n\" semin\"/\" sme\", predicting \"ar\" and forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3249)\nA/1/2083\n\" bul\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2083)\nA/1/2756\nStart-of-word \"Imag\"/\"fasc\", predicting \"ination\" or forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2756)\nA/1/1547\n\" cond\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1547)\nA/1/1482\n\" spir\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1482)\nA/1/2155\n\" sem\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2155)\nA/1/503\n\"cre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-503)\nA/1/821\n\" Domin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-821)\nA/1/2478\n\" ple\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2478)\nA/1/1396\n\" pen\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1396)\nA/1/3418\n\" clim\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3418)\nA/1/2944\n\" solic\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2944)\nA/1/1226\n\" comm\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1226)\nA/1/860\n\"aph\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-860)\nA/1/2887\nFirst halves of \"-ulate\" words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2887)\nA/1/2678\n\" ins\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2678)\nA/1/921\nWord beginnings \" Adv\"/\" suff\"/\" coinc\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-921)\nA/1/2330\n\" att\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2330)\nA/1/2111\n\" sin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2111)\nA/1/858\n\" stabil\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-858)\nA/1/1810\n\" fin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1810)\nA/1/2236\n\" hom\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2236)\nA/1/3759\n\" throm\" in medical contexts, predicting \"bo-\" tokens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3759)\nA/1/1596\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1596)\nA/1/3038\n\" organ\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3038)\nA/1/305\n\" prima\" (in \"prima facie\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-305)\nA/1/700\n\"nih\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-700)\nA/1/3068\n\" syn\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3068)\nCluster #401\nA/0/241\nThe feature fires when it sees words related to photography and imaging such as photo, micrograph, videotape, photogenic, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-241)\nA/1/120\n\" mon\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-120)\nA/1/624\nPrefixes like \" aut\" and \" Phot\", predicting \"ograph\" and similar suffixes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-624)\nA/1/772\n\" immun\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-772)\nA/1/1025\n\" hyp\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1025)\nA/1/1743\n\" psych\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1743)\nA/1/1973\n\" bi\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1973)\nA/1/2668\n\" hist\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2668)\nA/1/3603\n\" cert\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3603)\nA/1/3934\n\" radi\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3934)\nCluster #402\nA/0/225\nThis feature attends to words with the beginning prefix \"macro\" and stem \"mol\" such as macromolecule, macromolecules, and macromolecular.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-225)\nA/1/378\n\" phosph\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-378)\nA/1/669\n\"ozyg\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-669)\nA/1/923\n\" gran\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-923)\nA/1/1363\n\" ur\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1363)\nA/1/1624\n\" lip\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1624)\nA/1/1739\n\" cere\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1739)\nA/1/1891\n\" lymph\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1891)\nA/1/3642\nThis feature fires when it sees medical terms ending in -itis, -emia, and -osis, as well as the word mellitus.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3642)\nA/1/2724\n\"oph\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2724)\nCluster #403\nA/1/445\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-445)\nA/1/510\n\" flu\", in the context of bio or medical sciencey stuff, e.g. \"fluorescence\". Also fires on some other bio sciencey prefixes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-510)\nA/1/656\nMedical terms for cancers and diseases of the liver\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-656)\nA/1/2659\n\"orph\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2659)\nA/1/3357\n\"roph\" mostly in \"spectrophoto-\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3357)\nA/1/2848\n\"om\"/\"rim\" largely in cell biology\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2848)\nA/1/3325\n\"olog\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3325)\nA/1/854\n\" patri\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-854)\nA/1/1924\n\" tyr\"/\" handic\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1924)\nA/1/1578\n\" polit\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1578)\nA/1/1307\n\" ser\" prefix in seroepidemiological words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1307)\nA/1/2487\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2487)\nCluster #394\nA/1/535\n\" electroly\" or \"analy\", predicting \"tic\" or similar, as in \"analytically\" or \"electrolysis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-535)\nA/1/1562\n\" parame\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1562)\nA/1/3353\n\" plas\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3353)\nCluster #400\nA/0/43\nThis feature appears to fire for words containing the substring \"min\" such as dominant, dominik, domin, dominates. This includes variations like domination, mination, domination as well as words like minimum, minimums.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-43)\nA/0/408\nThis feature attends to words that are often found past tense or past participle verb forms.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-408)\nA/1/84\n\" altern\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-84)\nA/1/897\nThe feature attends to words ending with \"ation\" such as evaporation, proportional, optical or transportation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-897)\nA/1/1032\nWord prefixes like \" incompet\" and \" independ\" that predict endings like \"ent\" or \"ence\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1032)\nA/1/1044\n\" expl\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1044)\nA/1/1107\n\" ped\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1107)\nA/1/1139\n\" domin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1139)\nA/1/1154\n\" illumin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1154)\nA/1/1774\n\" Contin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1774)\nA/1/1967\n\" interle\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1967)\nA/1/2456\n\" Rel\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2456)\nA/1/2457\n\"rom\" in the middle of scientific/mechanical words, esp. forms of \"macromolecule\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2457)\nA/1/2780\n\" lum\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2780)\nA/1/2928\n\" reass\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2928)\nA/1/3097\n\"sor\" in word beginning \" adsorb\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3097)\nA/1/3323\nMostly \" quant\", predicting forms of \"itate\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3323)\nA/1/3508\n\" neg\" / \"eg\" etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3508)\nA/1/3742\n\" substit\" and similar word prefixes, predicting \"ute\" and word forms\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3742)\nA/1/3997\nWord beginnings like \" indul\" and \" diver\" predicting suffixes like \"-ge\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3997)\nA/1/4019\nMid-word tokens after the letter \"p\", esp. \"lic\" in \"complicate\"/\"implicate\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4019)\nA/1/4058\n\" cran\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4058)\nA/1/2469\n\" dispar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2469)\nA/1/721\nThe feature fires on words containing \"ant\" and forms of that root such as anti, anterior, antisera, etc. In particular it attends to medical and biological terms containing these roots.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-721)\nA/1/2415\n\" ann\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2415)\nA/1/2808\n\" tend\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2808)\nA/1/3055\n\" contra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3055)\nA/1/571\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-571)\nCluster #399\nA/1/60\n\" mot\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-60)\nA/1/957\nMostly \" tub\" with some \" tum\", \" trib\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-957)\nA/1/1823\n\"ag\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1823)\nA/1/2158\n\"ult\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2158)\nA/1/2505\nThis feature responds to words related to light and vision, like \"photoluminescence\", \"photoreceptor\", and \"photodetector\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2505)\nA/1/2436\nMid-word tokens after \"un-\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2436)\nA/1/1675\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1675)\nA/1/316\n\" judic\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-316)\nA/1/1549\n\" conce\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1549)\nA/1/2710\nMid-word tokens after \"dis\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2710)\nA/1/754\nUsually \"pe\" in \"Appeal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-754)\nA/1/3309\n\"Abbrev\" in Abbreviations\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3309)\nA/1/2137\nMid-word tokens esp. after \"emb\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2137)\nA/1/3859\n\" lit\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3859)\nA/1/3882\n\" Prom\"/\"Prom\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3882)\nA/1/2170\nLetters after \"qu\" in a word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2170)\nA/1/4073\n\" mill\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4073)\nCluster #373\nA/1/392\n\" ban\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-392)\nA/1/606\n\" par\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-606)\nA/1/912\n\" tor\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-912)\nA/1/1314\n\" gar\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1314)\nA/1/467\n\" dem\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-467)\nCluster #370\nA/0/92\nThis feature attends to words ending in an 'b' sound.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-92)\nA/1/164\nTokens after a single lowercase letter in a word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-164)\nA/1/2162\nMid-word tokens in words starting with \"s\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2162)\nA/1/2874\n\"ocal\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2874)\nA/1/2588\n\" gle\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2588)\nA/1/3771\nThis feature attends to words related to introducing, producing or inducing something.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3771)\nA/1/114\n\" ther\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-114)\nA/1/295\n\" revol\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-295)\nA/1/2815\n\" gro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2815)\nA/1/106\n\" cro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-106)\nA/1/2429\n\" cra\"/\" scra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2429)\nA/1/76\n\" den\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-76)\nA/1/1073\n\" che\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1073)\nA/1/3021\n\" fil\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3021)\nA/1/2348\n\" gra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2348)\nA/1/3688\n\" mo\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3688)\nA/1/1777\n\" pe\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1777)\nA/1/2195\n\" ro\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2195)\nCluster #377\nA/1/81\n\" cre\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-81)\nA/1/83\n\" ve\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-83)\nA/1/523\n\" ra\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-523)\nA/1/2413\n\" ce\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2413)\nCluster #378\nA/0/355\nThis feature seems to attend to words that contain \"se\" or \"le\". This pattern holds across multiple languages, including English, Spanish, French, German, Italian, Portuguese, and Romanian. The feature fires most strongly for words where \"se\" or \"le\" is in the middle of the word.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-355)\nA/1/140\n\" lo\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-140)\nA/1/1188\n\" bo\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1188)\nA/1/1278\n\" te\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1278)\nA/1/2866\n\" le\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2866)\nA/1/3419\n\" se\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3419)\nA/1/3787\n\" ne\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3787)\nA/1/1993\n\" ma\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1993)\nA/1/3911\n\" bu\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3911)\nA/1/2476\n\" une\" or \" sa\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2476)\nA/1/2271\n\" Los\"/\"Los\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2271)\nCluster #182\nA/1/155\n\" su\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-155)\nA/1/1358\n\" Je\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1358)\nA/1/3883\n\" du\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3883)\nA/1/479\n\" de\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-479)\nA/1/1477\n\" av\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1477)\nCluster #193\nA/0/76\nThis feature seems to attend to Spanish and other Romance languages. Many of the high activations contain words like \"de\", \"la\", \"el\", etc. which are common in Spanish, as well as characters like á, é, í, etc. The feature also seems to fire on French, Italian, and Portuguese words.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-76)\nA/0/206\n\"la\" in french and spanish\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-206)\nA/1/597\nItalian or Spanish: Feminine articles \" una\"/\" la\" in Spanish and maybe Italian / other languages?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-597)\nA/1/1251\nSpanish: \" de\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1251)\nA/1/1965\nSpanish: Word beginnings?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1965)\nA/1/379\n\" di\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-379)\nCluster #208\nA/0/61\nThis feature seems to fire when it sees text in Portuguese or Brazilian Portuguese.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-61)\nA/0/487\nThe feature seems to fire for Spanish words ending in -aban, -aban, -aron, -aban, or -aban. This indicates it is picking up on the Spanish imperfect past tense endings for -ar verbs.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-487)\nA/1/30\nSpanish: \" que\", usually Spanish?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-30)\nA/1/434\n\" no\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-434)\nA/1/542\nSpanish: \" el\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-542)\nA/1/725\nSpanish: word endings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-725)\nA/1/944\nSpanish or Portuguese: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-944)\nA/1/1337\nSpanish: \" en\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1337)\nA/1/1997\nSpanish: Punctuation, esp. colons\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1997)\nA/1/2109\nPortuguese tokens, esp. \" prepar\", predicting \"ão\" and endings\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2109)\nA/1/2244\nSpanish: \" los\" and other articles?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2244)\nA/1/2331\nSpanish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2331)\nA/1/2502\nPortuguese: Prepositions/articles\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2502)\nA/1/2989\nPortuguese: Strong tokens with accented and distinguishing Portuguese characters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2989)\nA/1/3024\nPortuguese: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3024)\nA/1/3140\nSpanish or Portuguese: \" a\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3140)\nA/1/3656\nSpanish: \"reg\" and other mid-word tokens?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3656)\nA/1/3716\nSpanish: \"Yo\" / \"adie\" / ... ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3716)\nA/1/3836\nShort word fragments esp. just one capital letter\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3836)\nA/1/833\nSpanish: Space+letter tokens?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-833)\nCluster #199\nA/0/14\nThis feature appears to fire when French is present.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-14)\nA/0/44\nThis feature fires when it sees accented characters, like é, ã, ö, â, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-44)\nA/0/182\nThis feature seems to fire on apostrophes in French text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-182)\nA/0/255\nThis feature fires when it sees Spanish words related to government, legal, and business topics such as politics, legislation, petitions, solicitation, requests, and registration.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-255)\nA/1/150\n\" est\"/\" Est\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-150)\nA/1/436\nFrench: \" les\"/\" des\" (the/some)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-436)\nA/1/544\nFrench: Apostrophe as contraction\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-544)\nA/1/956\nSpanish or Portuguese: Some -du word beginnings?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-956)\nA/1/981\nFrench: é in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-981)\nA/1/991\nFrench: \" qui\" but only in the that sense, not who\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-991)\nA/1/1050\nFrench: predicts é|e|è\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1050)\nA/1/1168\nFrench: \" é\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1168)\nA/1/1288\nFrench: \"è\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1288)\nA/1/1892\nFrench: Prepositions?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1892)\nA/1/2016\nFrench: \" :\", and other punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2016)\nA/1/2042\nFrench: \" dé\"/\" ré\" prefix only\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2042)\nA/1/2146\nFrench: Names / proper nouns, perhaps in a bibliography?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2146)\nA/1/3002\nFrench: Prepositions about containing/linking\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3002)\nA/1/3008\nFrench: Predicts prepositions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3008)\nA/1/3207\n\"\\u00e2\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3207)\nA/1/3596\nFrench: \" à\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3596)\nCluster #209\nA/0/184\nThis feature seems to be activating on words that have diphthongs (vowel sounds that glide from one position in the mouth to another) in Italian and other Romance languages.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-184)\nA/1/2034\nItalian: Formal governmental (Not sure why it it's matching on a subset of the Italian words/characters though)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2034)\nA/1/2156\nSpanish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2156)\nA/1/2250\nItalian: elements of verb conjugation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2250)\nCluster #200\nA/0/455\nThis feature attends to text in Romanian.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-455)\nA/1/9\nRomanian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-9)\nA/1/632\nRomanian: \" \\u00ee\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-632)\nA/1/846\nAccented, esp. acute, characters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-846)\nA/1/3923\nLatin: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3923)\nA/1/734\nIndonesian/Malaysian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-734)\nCluster #211\nA/0/159\nThis feature appears to fire on words ending in ng (and similar sounds like nk, nd, etc.), with a preference for words of three or more syllables.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-159)\nA/1/1657\nTurkish, Basque, Hawaiian, Maori, and other Polynesian languages: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1657)\nA/1/1912\nIcelandic: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1912)\nA/1/623\nVietnamese & Romance Languages: Text near UTF-8 as latin1 mojibake\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-623)\nCluster #212\nA/0/36\nThis feature is attending to Slavic languages, mainly Polish and Czech.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-36)\nA/0/294\nThis feature attends to Hungarian words and some other Hungarian-like examples from nearby languages.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-294)\nA/1/1141\n\"\\u00e1\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1141)\nA/1/1200\n\",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1200)\nA/1/1627\nTurkish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1627)\nA/1/1772\nEstonian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1772)\nA/1/2201\nLithuanian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2201)\nA/1/2360\nSlovenian: Text or names?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2360)\nA/1/2651\nSlovenian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2651)\nA/1/2676\nHungarian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2676)\nA/1/2792\nSlovenian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2792)\nA/1/2801\nSlavic: \",\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2801)\nA/1/3348\nPolish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3348)\nA/1/3361\nLatvian: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3361)\nA/1/3537\nVietnamese: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3537)\nA/1/3648\nCzech: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3648)\nA/0/180\nThis feature attends to punctuation such as periods, quotation marks, newlines.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-180)\nA/1/3329\nFinnish: Word beginnings?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3329)\nA/0/162\nThis feature fires when it sees German words that start with an \"s\" followed by a consonant.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-162)\nCluster #205\nA/0/339\nThis feature seems to be firing for Finnish words and sentences.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-339)\nA/1/1052\nFinnish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1052)\nA/1/2570\nFinnish and German?: Consonants after umlauts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2570)\nA/1/2897\nFinnish: Umlauts (and some surrounding text?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2897)\nA/1/3413\nFinnish: tokens esp. with \"-ist\" or \"-ll\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3413)\nA/1/2153\nSwedish?: Letters or tokens with diacritics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2153)\nA/1/1112\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1112)\nA/1/917\nDutch: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-917)\nCluster #213\nA/0/469\nThis feature seems to fire on Germanic languages - specifically Scandinavian languages like Swedish, Danish, and Norwegian. It is attuned to grammatical features and common words in those languages.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-469)\nA/1/321\nSwedish: Punctuation\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-321)\nA/1/1373\nDanish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1373)\nA/1/2464\nSwedish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2464)\nA/1/3046\nDanish: Politics\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3046)\nCluster #196\nA/0/493\nThis feature seems to activate for German text, particularly German grammar such as noun declensions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-493)\nA/1/177\nGerman: \" der\", plus other articles/prepositions weakly\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-177)\nA/1/334\nGerman: Text tokens ending in \"e\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-334)\nA/1/358\nGerman: Consonants at start of word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-358)\nA/1/838\nGerman: Prepositions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-838)\nA/1/1695\n\" und\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1695)\nA/1/2410\nGerman:?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2410)\nA/1/2601\nGerman: Commas\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2601)\nA/1/2637\nGerman: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2637)\nA/1/2652\nGerman: Pronouns (\"ihre\", \"die\", \"dieser\")\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2652)\nA/1/2661\nGerman: Modal verbs\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2661)\nCluster #214\nA/1/1951\nDutch: \" wa\"/\" na\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1951)\nA/1/2065\nDutch: Some word beginnings?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2065)\nA/1/2899\nDanish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2899)\nA/1/2032\nSwedish: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2032)\nA/1/103\nLaTeX length names, particularly \"side\" in \"oddsidemargin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-103)\nA/1/2576\n\"ensuremath\" in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2576)\nCluster #116\nA/0/35\nThis feature fires for math expressions, in particular fractions displayed using LaTeX.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-35)\nA/0/82\nThis feature attends to mathematical symbols.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-82)\nA/1/267\nLaTeX mathbb font\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-267)\nA/1/427\nLaTeX operators with a number over another (frac, binom, dfrac)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-427)\nA/1/914\nLaTeX macro names esp. accepting one short argument\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-914)\nA/1/1136\nLaTeX environment \"begin\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1136)\nA/1/1889\nLaTeX text modifiers in equations (inc. mathfrak, sqrt, overline, boldsymbol)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1889)\nA/1/2226\nThis feature seems to fire on LaTeX formatting and math markup.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2226)\nA/1/2590\nThis feature fires on TeX/LaTeX formatting such as equations, labels, sections, tables, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2590)\nCluster #3\nA/0/175\nThis feature attends to mathematical symbols and expressions.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-175)\nA/0/368\nThis feature fires when it encounters LaTeX commands and math symbols.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-368)\nA/1/112\nCode: \"\\\" in paths or string escapes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-112)\nA/1/193\nLaTeX: Inline modifiers\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-193)\nA/1/222\nLaTeX: Predicting an operator?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-222)\nA/1/262\nCode: \"\\\" in RTF source code?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-262)\nA/1/363\nLaTeX: Predicting command with alternate tokens\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-363)\nA/1/848\nLaTeX: Punctuation clusters ending in \\, predicting a command\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-848)\nA/1/1123\nLaTeX: \" $\\\", predicting greek letters in formulae\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1123)\nA/1/1277\nLaTeX: Predicting command\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1277)\nA/1/1506\nLaTeX: Bracket and \"\\\", predicting command (often formatting)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1506)\nA/1/1911\nLaTeX: \"(\\\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1911)\nA/1/1917\nLaTeX: Superscripted or subscripted backslash, predicting macros in superscripts/subscripts'\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1917)\nA/1/2313\nLaTeX: Predicting escaped characters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2313)\nA/1/2737\nLaTeX: \"\\\" and \" \\\" - left/right delimiters, etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2737)\nA/1/3205\nLaTeX: \" $$\\\\\\\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3205)\nA/1/3471\nLaTeX: \"^{\\\" and \"_{\\\", introducing superscripted/subscripted commands\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3471)\nA/1/3517\nBack Slash in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3517)\nA/1/3853\nLaTeX: Predicting equation symbols\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3853)\nA/1/3891\nLaTeX: \"{\\\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3891)\nA/1/3894\nLaTeX: Subscript with escape\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3894)\nCluster #109\nA/0/173\nThis feature fires for LaTeX markup, specifically for commands that start an environment or matrix.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-173)\nA/1/529\nLaTeX font setting, primarily \\rm\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-529)\nA/1/3789\n\"usepackage\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3789)\nCluster #77\nA/0/158\nThe feature attends to mathematical notation and formulas.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-158)\nA/0/205\nThis feature fires on mathematical notation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-205)\nA/1/108\nLaTeX contents of curly-braced superscripts or subscripts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-108)\nA/1/356\nLaTeX measurements (probably units)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-356)\nA/1/455\n\"aligned\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-455)\nA/1/2452\nContents, esp. digits, of curly braces in superscripts or subscripts, predicting close braces\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2452)\nA/1/3340\nLaTeX letters in \\mathcal and a few other fonts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3340)\nA/1/3552\nNumbers after open curly braces in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3552)\nA/1/1798\nCode: Letters in backslash escapes (n in \\n etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1798)\nA/1/2927\nLaTeX fraction numerators\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2927)\nA/1/3384\n\"69\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3384)\nA/1/723\nLaTeX: macro names for spacing like \"qquad\" or dots like \"ldots\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-723)\nCluster #88\nA/0/93\nThis feature attends to text written in LaTeX.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-93)\nA/1/633\n\"}\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-633)\nA/1/840\n\" \\\\\\\\\\\\\\\"/\"\\\\\\\\\\\\\\\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-840)\nA/1/1186\nHexadecimal digits in \\x escapes\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1186)\nCluster #98\nA/0/185\nThis feature appears to attend to mathematical equations with numbers, particularly those with the value of 1 or 2.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-185)\nA/1/788\nNumbers after \"=\" or similar operators, e.g. \\approx\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-788)\nA/1/2401\nLaTeX exponents (after \"^\"), esp. 2\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2401)\nA/1/2501\nNumbers after \"_\" (subscripts), often predicting superscripts\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2501)\nCluster #68\nA/1/1285\nPhysics notation, esp. near square brackets (hints of relativity)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1285)\nA/1/2221\nLaTeX: numbers in matrices/tables/etc.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2221)\nA/1/2785\nThis feature attends to mathematical notation with formulas and equations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2785)\nCluster #117\nA/0/111\nThis feature fires when it sees mathematical notation, specifically symbols used in calculus and algebra.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-111)\nA/0/161\nThis feature attends to mathematical formulas written in LaTeX.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-161)\nA/0/499\nThis feature fires for mathematical variables.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-499)\nA/1/137\nLaTeX big delimiters\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-137)\nA/1/189\nLaTeX: \"right\" command\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-189)\nA/1/1035\nGreek letters in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1035)\nA/1/1576\nGreek letter macros in LaTeX math mode\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1576)\nA/1/1829\nLaTeX: short, mostly single-letter variables\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1829)\nA/1/1929\nThis feature fires when it sees mathematical equations and LaTeX-style notation.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1929)\nA/1/2056\nLaTeX trigonometric functions\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2056)\nA/1/2512\nLaTeX left delimiter opening\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2512)\nA/1/2750\n\\to in LaTeX \nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2750)\nA/1/2800\nLaTeX operators (sum, prod, int, etc.)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2800)\nA/1/2920\nLaTeX Math (general)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2920)\nA/1/3057\nFeature for predicting \"_{\" and similar in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3057)\nA/1/3333\nSingle-letter tokens/variables in LaTeX math mode\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3333)\nA/1/3574\nGreek letters in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3574)\nA/1/3752\nOne-letter variables, often subscripts, in LaTeX math (predicting superscripts)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3752)\nA/1/3940\nSpace + capital letter in LaTeX often a variable after a macro like \"\\in\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3940)\nA/1/4044\n\"partial\" in the LaTeX macro \"\\partial\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4044)\nA/1/697\nNumbers in parentheses, often comma-separated tuples, in LaTeX math mode\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-697)\nA/1/2295\nSingle-digit numbers in LaTeX math\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2295)\nA/0/144\nThis feature attends to mathematical formulas. It activates on variables, mathematical constants and operators, as well as formatting and punctuation commonly used in equations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-144)\nA/1/2864\nThis feature fires in response to mathematical formulae, particularly those containing mathematical symbols such as Greek letters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2864)\nA/1/596\nLatin-alphabet identifiers in LaTeX\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-596)\nA/1/3258\nLaTeX: variables in parenthesized expressions or pairs in inline math mode\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3258)\nA/0/167\nThis feature appears to fire for code examples and programming language keywords, especially in C/C++.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-167)\nCluster #75\nA/0/275\nThis feature attends to capital letter A at the beginning of sentences.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-275)\nA/0/325\nThis feature fires when it sees a single capital letter followed by a period, most often at the beginning of a sentence.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-325)\nA/1/654\n\"S\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-654)\nA/1/806\nThis feature attends to acronyms, initialisms and abbreviations in the text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-806)\nA/1/1207\n'g' in 'e.g.', predicts the next period\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1207)\nA/1/1773\n\" A\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1773)\nA/1/1998\n\" A\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1998)\nA/1/2243\nThis feature attends to US Court case citations.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2243)\nA/1/3625\n\"A\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3625)\nA/1/4014\n\"L\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4014)\nCluster #121\nA/1/784\n\" X\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-784)\nA/1/1429\n\" N\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1429)\nA/1/2125\n\" D\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2125)\nA/1/2776\n\" V\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2776)\nCluster #70\nA/0/213\nThis feature attends to words that contain the letter \"b\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-213)\nA/1/300\n\" d\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-300)\nA/1/438\n\" e\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-438)\nA/1/476\n\" w\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-476)\nA/1/530\n\" k\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-530)\nA/1/738\n\" m\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-738)\nA/1/825\n\" h\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-825)\nA/1/910\n\" p\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-910)\nA/1/1079\n\" r\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1079)\nA/1/1779\n\" j\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1779)\nA/1/1790\n\"f\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1790)\nA/1/2108\n\" b\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2108)\nA/1/2265\n\" t\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2265)\nA/1/2332\n\" s\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2332)\nA/1/2335\n\" l\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2335)\nA/1/2667\n\" g\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2667)\nA/1/3371\n\" c\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3371)\nA/1/3579\n\" v\" usually beginning a word\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3579)\nA/1/3730\n\" n\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3730)\nA/1/3993\n\" o\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3993)\nCluster #120\nA/0/26\nThis feature fires when it sees initials, especially two initials with a period between them. It attends to proper nouns and names in abbreviation format.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-26)\nA/0/137\nThis feature fires on the letter D in text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-137)\nA/0/296\nThis feature fires in response to proper nouns that start with \"F\" or some other capital letters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-296)\nA/0/395\nThis feature is firing for instances of \"X.P\" and \"P.X\" where \"X\" represents uppercase Roman numerals and \"P\" represents single digit page numbers.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-395)\nA/0/479\nThe feature fires when it sees capital M followed by a capital letter.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-479)\nA/1/11\n\" M\" esp. shortly after <EOT>\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-11)\nA/1/323\n\" Z\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-323)\nA/1/1057\n\" P\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1057)\nA/1/1871\n\" T\" in titles/names/English\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1871)\nA/1/3052\n\" C\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3052)\nA/1/3107\n\" Y\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3107)\nA/1/3159\n\" J\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3159)\nA/1/3289\n\" G\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3289)\nA/1/3406\n\" W\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3406)\nA/1/3408\n\" F\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3408)\nA/1/3428\n\" L\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3428)\nA/1/3939\n\" R\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3939)\nA/1/732\n\" H\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-732)\nA/1/1740\n\" K\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1740)\nA/1/1786\n\" B\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1786)\nA/1/748\nThis feature fires for character names that start with a soft C, D, F, J, K, M, N, P, R, T, V, W, Z sound.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-748)\nA/1/1010\n\" S\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1010)\nA/1/1718\n\"C\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1718)\nA/0/350\nThis feature attends to capital S followed by 4-9 capital letters.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-350)\nCluster #93\nA/0/278\nThis feature fires for names and words starting with the letter E.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-278)\nA/1/1377\n\" O\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1377)\nA/1/3351\n\" U\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3351)\nA/1/4008\n\" E\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4008)\nA/1/3528\n\"B\" esp. after \"-\" or \"/\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3528)\nCluster #1\nA/0/397\nThis feature attends to Chinese sentences and phrases.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-397)\nA/1/374\nChinese: Technical\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-374)\nA/1/900\nChinese: Simplified\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-900)\nA/1/2000\nChinese: Obscure characters / dictionaries / transliterated names\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2000)\nA/1/2030\nJapanese: Numbers (dates and laws?)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2030)\nCluster #2\nA/0/304\nThis feature fires on Japanese text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-304)\nA/1/698\nJapanese: Period, end-of-sentence character\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-698)\nA/1/1260\nChinese: Negations / \"but\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1260)\nA/1/2730\nJapanese: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2730)\nCluster #8\nA/1/1671\nJapanese and Chinese: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1671)\nA/1/2239\nJapanese: \"そ\"?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2239)\nA/1/3655\nJapanese, Chinese, and Korean: ?\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3655)\nCluster #5\nA/1/2602\nJapanese: The prefix the \"suru\" verb that means \"to do\", when used in a bunch of contexts, but _without_ the conjugation (ie, positive/negative, future/past, volitional/not, want-to/not, etc)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2602)\nA/1/2986\nJapanese: The \"です desu\" copula that means \"is\". It represents \"X is Y\" in \"X Y desu\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2986)\nA/1/3112\nJapanese: Suffix that noun-ize a verb, similar to \"ing\" suffix in english\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3112)\nA/1/3605\nJapanese: A form of the \"to do\" verb prefix. Usually passive tense, occasionally \"can/able to do\", and sometimes \"is doing\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3605)\nA/1/3927\nJapanese: Hiragana \"e\" character, predicting things likely to come after that\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3927)\nA/1/3965\nJapanese: \"な\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3965)\nCluster #9\nA/0/343\nThis feature fires when it sees Japanese text.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-343)\nA/1/63\nJapanese: Hiragana particle that follows a noun, representing it was a direct or indirect object\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-63)\nA/1/380\nJapanese: Seems to be the \"to X\" form of the verb, represented in an impolite form, sometimes used in conjunction with suffixes that \"noun-ize\" the verb\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-380)\nA/1/421\nJapanese: Specialized - game/IRC chats\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-421)\nA/1/964\nJapanese: Katakana characters representing foreign-loan words\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-964)\nA/1/2850\nJapanese: Hiragana ha/wa character, that is particle representing the phrase before it is a \"topic\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2850)\nA/1/3829\nJapanese: Hiragana \"no\" that represents possessive conjunction, similar to \"apostrophe-s\" in English\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3829)\nCluster #22\nA/0/267\nThis feature fires on words that indicate a transition in a sentence, such as prepositions like \"to\", \"in\", \"for\", etc. It seems to be particularly attentive to the word \"to\".\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-267)\nA/0/331\nThis feature attends to the use of the preposition ``to'' in a variety of functional contexts.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-331)\nA/1/297\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-297)\nA/1/390\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-390)\nA/1/975\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-975)\nA/1/990\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-990)\nA/1/1541\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1541)\nA/1/2169\n\" To\"\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2169)\nA/1/2439\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2439)\nA/1/2543\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2543)\nA/1/3613\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3613)\nA/1/3854\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3854)\nA/1/3484\n\"to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3484)\nA/1/2240\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2240)\nCluster #23\nA/0/226\nThe feature fires when the word \"to\" is used to indicate an infinitive verb.\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-226)\nA/1/1262\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1262)\nA/1/1364\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1364)\nA/1/1802\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1802)\nA/1/1832\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1832)\nA/1/2063\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2063)\nA/1/2191\n\" to\" in legal\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2191)\nA/1/2361\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2361)\nA/1/2653\n[Ultralow density cluster]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2653)\nA/1/2666\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2666)\nA/1/2688\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2688)\nA/1/2960\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2960)\nA/1/3018\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3018)\nA/1/3297\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3297)\nA/1/3778\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3778)\nA/1/3805\n\" to\" in infinitive phrases (verb \" to\" verb)\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3805)\nA/1/4011\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4011)\nA/1/2172\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2172)\nA/1/3956\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3956)\nA/1/1047\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1047)\nA/1/905\n\" to\" in [context]\nZoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-905)\nscroll zoom homezoom inzoom out\nA/1/207A/1/300A/1/334A/1/390A/1/593A/1/1015A/1/1336A/1/1393A/1/1411A/1/1506A/1/1524A/1/1648A/1/1851A/1/2000A/1/2030A/1/2095A/1/2361A/1/2482A/1/2625A/1/2645A/1/2651A/1/2782A/1/3041A/1/3042A/1/3205A/1/3333A/1/3471A/1/3474A/1/3579A/1/3875A/1/3965A/1/3967A/1/4082\n### Summary of Results\n  * Sparse Autoencoders extract relatively monosemantic features. We provide four different lines of evidence: [detailed investigations](https://transformer-circuits.pub/2023/monosemantic-features/#feature-analysis) for a few features firing in specific contexts for which we can construct computational proxies, [human analysis](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-manual) for a large random sample of features, [automated interpretability analysis of activations](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-auto-acts) for all the features learned by the autoencoder, and finally [automated interpretability analysis of logit weights](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-auto-logits) for all the features. Moreover, the last three analyses show that most learned features are interpretable. While we do not claim that our interpretations catch all aspects of features' behaviors, by constructing metrics of interpretability consistently for features and neurons, we quantitatively show their relative interpretability.\n  * Sparse autoencoders produce interpretable features that are effectively invisible in the neuron basis. We find features (e.g., [one firing on Hebrew script](https://transformer-circuits.pub/2023/monosemantic-features/#feature-hebrew)) which are not active in any of the top dataset examples for any of the neurons.\n  * Sparse autoencoder features can be used to intervene on and steer transformer generation. For example, activating the base64 feature we study causes the model to generate base64 text, and activating the Arabic script feature we study produces Arabic text. (See [discussion](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-about-model) of pinned feature sampling in [Global Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis).)\n  * Sparse autoencoders produce relatively universal features. Sparse autoencoders applied to different transformer language models produce mostly similar features, more similar to one another than they are to their own model's neurons. (See [Universality](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-universality))\n  * Features appear to \"split\" as we increase autoencoder size. When we gradually increase the width of the autoencoder from 512 (the number of neurons) to over 131,000 (256×), we find features which naturally fit together into families. For example, one base64 feature in a small dictionary [splits into three](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs-2), with more subtle and yet still interpretable roles, in a larger dictionary. The different size autoencoders offer different \"resolutions\" for understanding the same object. (See [Feature Splitting](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-splitting).)\n  * Just 512 neurons can represent tens of thousands of features. Despite the MLP layer being very small, we continue to find new features as we scale the sparse autoencoder.\n  * Features connect in \"finite-state automata\"-like systems that implement complex behaviors. For example, we find features that work together to generate valid HTML. (See [\"Finite State Automata\"](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-fsa).)\n\n  \n  \n  \n\n* * *\n  \n  \n\n## [Problem Setup](https://transformer-circuits.pub/2023/monosemantic-features/#problem-setup)\nA key challenge to our agenda of reverse engineering neural networks is the curse of dimensionality: as we study ever-larger models, the volume of the latent space representing the model's internal state that we need to interpret grows exponentially. We do not currently see a way to understand, search or enumerate such a space unless it can be decomposed into independent components, each of which we can understand on its own.\nIn certain limited cases, it is possible to side step these issues by rewriting neural networks in ways that don't make reference to certain hidden states. For example, in A Mathematical Framework for Transformer Circuits , we were able to analyze a one-layer attention-only network without addressing this problem. But this approach becomes impossible if we consider even a simple standard one-layer transformer with an MLP layer with a ReLU activation function. Understanding such a model requires us to have a way to decompose the MLP layer.\nIn some sense, this is the simplest language model we profoundly don't understand. And so it makes a natural target for our paper. We aim to take its MLP activations – the activations we can't avoid needing to decompose – and decompose them into \"features\":\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nCrucially, we decompose into more features than there are neurons. This is because we believe that the MLP layer likely uses superposition to represent more features than it has neurons (and correspondingly do more useful non-linear work!). In fact, in our largest experiments we'll expand to have 256 times more features than neurons, although we'll primarily focus on a more modest 8× expansion.  \n|   | Transformer  | Sparse Autoencoder  |  \n| --- | --- | --- |  \n| Layers  | 1 Attention Block  \n1 MLP Block (ReLU)  | 1 ReLU (up)  \n1 Linear (down)  |  \n| MLP Size  | 512  | 512 (1×) – 131,072 (256×)  |  \n| Dataset  | The Pile  \n(100 billion tokens)  | Transformer MLP Activations   \n(8 billion samples)  |  \n| Loss  | Autoregressive Log-Likelihood  | L2 reconstruction  \n+ L1 on hidden layer activation  |  \nIn the following subsections, we will motivate this setup at more length. Additionally, a more detailed discussion of the architectural details and training of these models can be found in the appendix.\n### [Features as a Decomposition](https://transformer-circuits.pub/2023/monosemantic-features/#problem-setup-decomposition)\nThere is significant empirical evidence suggesting that neural networks have interpretable linear directions in activation space. This includes classic work by Mikolov et al.  investigating \"vector arithmetic\" in word embeddings (but see ) and similar results in other latent spaces (e.g. ). There is also a large body of work investigating interpretable neurons, which are just basis dimensions (e.g. in RNNs ; in CNNs ; in GANs ; but see ). A longer review and discussion of this work can be found in the [Motivation section](https://transformer-circuits.pub/2022/toy_model/index.html#motivation) of Toy Models , although it doesn't cover more recent work (e.g. ).We'd particularly highlight an exciting exchange between Li et al. and Nanda et al. : Li et al. found an apparent counterexample where features were not represented as directions, which was then resolved by Nanda finding an alternative interpretation in which features were directions.\nIf linear directions are interpretable, it's natural to think there's some \"basic set\" of meaningful directions which more complex directions can be created from. We call these directions features, and they're what we'd like to decompose models into. Sometimes, by happy circumstances, individual neurons appear to be these basic interpretable units (see examples above). But quite often, this isn't the case. \nInstead, we decompose the activation vector \\mathbf{x}^j as a combination of more general features which can be any direction:\n\\mathbf{x}^j \\approx \\mathbf{b} + \\sum_i f_i(\\mathbf{x}^j) \\mathbf{d}_i **(1)**\nwhere \\mathbf{x}^j is the activation vector of length d_{\\rm MLP} for datapoint j, f_i(\\mathbf{x}^j) is the activation of feature i, each \\mathbf{d}_i is a unit vector in activation space we call the direction of feature i, and \\mathbf{b} is a bias.While not the focus of this paper, we could also imagine decomposing other portions of the model’s latent state or activations in this way. For instance, we could apply this decomposition to the residual stream (see e.g. ), or to the keys and queries of attention heads, or to their outputs. Note that this decomposition is not new: it is just a linear matrix factorization of the kind commonly employed in dictionary learning.\nIn our sparse autoencoder setup, the feature activations are the output of the encoder\nf_i(x) = \\text{ReLU}( W_e (\\mathbf{x} - \\mathbf{b}_d) + \\mathbf{b}_e )_i,\nwhere W_e is the weight matrix of the encoder and \\mathbf{b}_d, \\mathbf{b}_e are a pre-encoder and an encoder bias. The feature directions are the columns of the decoder weight matrix W_d. (See [appendix](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder) for full notation.) \nIf such a sparse decomposition exists, it raises an important question: are models in some fundamental sense composed of features or are features just a convenient post-hoc description? In this paper, we take an agnostic position, though our [results](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-universality) on feature universality suggest that features have some existence beyond individual models.\n#### Superposition Hypothesis\nTo see how this decomposition relates to superposition, recall that the superposition hypothesis postulates that neural networks “want to represent more features than they have neurons”. We think this happens via a kind of “noisy simulation”, where small neural networks exploit feature sparsity and properties of high-dimensional spaces to approximately simulate much larger much sparser neural networks .\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nA consequence of this is that we should expect the feature directions to form an overcomplete basis. That is, our decomposition should have more directions \\mathbf{d}_i than neurons. Moreover, the feature activations should be sparse, because sparsity is what enables this kind of noisy simulation. This is mathematically identical to the classic problem of dictionary learning.\n### [What makes a good decomposition?](https://transformer-circuits.pub/2023/monosemantic-features/#problem-setup-good-decomposition)\nSuppose that a dictionary exists such that the MLP activation of each datapoint is in fact well approximated by a sparse weighted sum of features as in equation 1. That decomposition will be useful for interpreting the neural network if:\n  1. We can interpret the conditions under which each feature is active. That is, we have a description of which datapoints j cause the feature to activate (i.e. for which f_i(\\mathbf{x}^j) is large) that makes sense on a diverse dataset (including potentially synthetic or off-distribution examples meant to test the hypothesis).This property is closely related to the desiderata of Causality, Generality, and Purity discussed in Cammarata et al. , and those provide an example of how we might make this property concrete in a specific instance.\n  2. We can interpret the downstream effects of each feature, i.e. the effect of changing the value of f_i on subsequent layers. This should be consistent with the interpretation in (1).\n  3. The features explain a significant portion of the functionality of the MLP layer (as measured by the loss; see [How much of the model does our interpretation explain](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-how-much)?).\n\n\nA feature decomposition satisfying these criteria would allow us to:\n  1. Determine the contribution of a feature to the layer’s output, and the next layer’s activations, on a specific example.\n  2. Monitor the network for the activation of a specific feature (see e.g. speculation about [safety-relevant features](https://transformer-circuits.pub/2023/july-update/index.html#safety-features)).\n  3. Change the behavior of the network in predictable ways by changing the values of some features. In multilayer models this could look like predictably influencing one layer by changing the feature activations in an earlier layer. \n  4. Demonstrate that the network has learned certain properties of the data.\n  5. Demonstrate that the network is using a given property of the data in producing its output on a specific example.This is similar in spirit to the evidence provided by influence functions.\n  6. Design inputs meant to activate a given feature and elicit certain outputs.\n\n\nOf course, decomposing models into components is just the beginning of the work of mechanistic interpretability! It provides a foothold on the inner workings of models, allowing us to start in earnest on the task of unraveling circuits and building a larger-scale understanding of models.\n### [Why not use architectural approaches?](https://transformer-circuits.pub/2023/monosemantic-features/#why-not-architectures)\nIn [Toy Models of Superposition](https://transformer-circuits.pub/2022/toy_model/index.html) , we highlighted several different approaches to solving superposition. One of those approaches was to engineer models to simply not have superposition in the first place.\nInitially, we thought that this might be possible but come with a large performance hit (i.e. produce models with greater loss). Even if this performance hit had been too large to use in practice for real models, we felt that success at creating monosemantic models would have been very useful for research, and in a lot of ways this felt like the \"cleanest\" approach for downstream analysis.\nUnfortunately, having spent a significant amount of time investigating this approach, we have ultimately concluded that it is more fundamentally non-viable.\nIn particular, we made several attempts to induce activation sparsity during training to produce models without superposition, even to the point of training models with 1-hot activations. This indeed eliminates superposition, but it fails to result in cleanly-interpretable neurons! Specifically, we found that individual neurons can be polysemantic even in the absence of superposition. This is because in many cases models achieve lower loss by representing multiple features ambiguously (in a polysemantic neuron) than by representing a single feature unambiguously and ignoring the others.\nTo understand this, consider a toy model with a single neuron trained on a dataset with four mutually-exclusive features (A/B/C/D), each of which makes a distinct (correct) prediction for the next token, labeled in the same fashion. Further suppose that this neuron’s output is binary: it either fires or it doesn’t. When it fires, it produces an output vector representing the probabilities of the different possible next tokens.\nWe can calculate the cross-entropy loss achieved by this model in a few cases:\n  1. Suppose the neuron only fires on feature A, and correctly predicts token A when it does. The model ignores all of the other features, predicting a uniform distribution over tokens B/C/D when feature A is not present. In this case the loss is \\frac{3}{4}\\ln 3 \\approx 0.8.\n  2. Instead suppose that the neuron fires on both features A and B, predicting a uniform distribution over the A and B tokens. When the A and B features are not present, the model predicts a uniform distribution over the C and D tokens. In this case the loss is \\ln 2 \\approx 0.7.\n\n\nBecause the loss is lower in case (2) than in case (1), the model achieves better performance by making its sole neuron polysemantic, even though there is no superposition. \nThis example might initially seem uninteresting because it only involves one neuron, but it actually points at a general issue with highly sparse networks. If we push activation sparsity to its limit, only a single neuron will activate at a time. We can now consider that single neuron and the cases where it fires. As seen earlier, it can still be advantageous for that neuron to be polysemantic. \nBased on this reasoning, and the results of our experiments, we believe that models trained on cross-entropy loss will generally prefer to represent more features polysemantically than to represent fewer \"true features\" monosemantically, even in cases where sparsity constraints make superposition impossible.\nModels trained on other loss functions do not necessarily suffer this problem. For instance, models trained under mean squared error loss (MSE) may achieve the same loss for both polysemantic and monosemantic representations (e.g. ), and for some feature importance curves they may actively prefer the monosemantic solution . But language models are not trained with MSE, and so we do not believe architectural changes can be used to create fully monosemantic language models.\nNote, however, that in learning to decompose models post-training we do use an MSE loss (between the activations and their representation in terms of the dictionary), so sparsity can inhibit superposition from forming in the learned dictionary. (Otherwise, we might have superposition \"all the way down.\")\n### [Using Sparse Autoencoders To Find Good Decompositions](https://transformer-circuits.pub/2023/monosemantic-features/#setup-autoencoder-motivation)\nThere is a long-standing hypothesis that many natural latent variables in the world are sparse (see , section \"Why Sparseness?\"). Although we have a limited understanding of the features that exist in language models, the examples we do have (e.g. ) are suggestive of highly sparse variables. Our work on Toy Models of Superposition shows that a large set of sparse features could be represented in terms of directions in a lower dimensional space.\nFor this reason, we seek a decomposition which is sparse and overcomplete. This is essentially the problem of [sparse dictionary learning](https://en.wikipedia.org/wiki/Sparse_dictionary_learning). (Note that, since we're only asking for a sparse decomposition of the activations, we make no use of the downstream effects of \\mathbf{d}_i on the model's output. This means we'll be able to use those downstream effects in later sections as a kind of validation on the features found.)\nIt's important to understand why making the problem overcomplete – which might initially sound like a trivial change – actually makes this setting very different from similar approaches seeking sparse disentanglement in the literature. It's closely connected to why dictionary learning is such a non-trivial operation; in fact, as we'll see, it's actually kind of miraculous that this is possible at all. At the heart of dictionary learning is an inner problem of computing the feature activations f_i(\\mathbf{x}) for each datapoint \\mathbf{x}, given the feature directions \\mathbf{d}_i. On its surface, this problem may seem impossible: we're asking to determine a high-dimensional vector from a low-dimensional projection. Put another way, we're trying to invert a very rectangular matrix. The only thing which makes it possible is that we are looking for a high-dimensional vector that is sparse! This is the famous and well-studied problem of [compressed sensing](https://en.wikipedia.org/wiki/Compressed_sensing), which is NP-hard in its exact form. It is possible to store high-dimensional sparse structure in lower-dimensional spaces, but recovering it is hard.\nDespite its difficulty, there are a host of sophisticated methods for dictionary learning (e.g. ). These methods typically involve optimizing a relaxed problem or doing a greedy search. We tried several of these classic methods, but ultimately decided to focus on a simpler sparse autoencoder approximation of dictionary learning (similar to Sharkey, et al. ). This was for two reasons. First, a sparse autoencoder can readily scale to very large datasets, which we believe is necessary to characterize the features present in a model trained on a large and diverse corpus.For the model discussed in this paper, we trained the autoencoder on 8 billion datapoints. Secondly, we have a concern that iterative dictionary learning methods might be [\"too strong\"](https://transformer-circuits.pub/2023/may-update/index.html#dictionary-worries), in the sense of being able to recover features from the activations which the model itself cannot access. Exact compressed sensing is NP-hard, which the neural network certainly isn't doing. By contrast, a sparse autoencoder is very similar in architecture to the MLP layers in language models, and so should be similarly powerful in its ability to recover features from superposition.\n### [Sparse Autoencoder Setup](https://transformer-circuits.pub/2023/monosemantic-features/#setup-autoencoder)\nWe briefly overview the architecture and training of our sparse autoencoder here, and provide further details in [Basic Autoencoder Training](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder). Our sparse autoencoder is a model with a bias at the input, a linear layer with bias and ReLU for the encoder, and then another linear layer and bias for the decoder. In toy models we found that the bias terms were quite important to the autoencoder’s performance.We tie the biases applied in the input and output, so the result is equivalent to subtracting a fixed bias from all activations and then using an autoencoder whose only bias is before the encoder activation.\nWe train this autoencoder using the Adam optimizer to reconstruct the MLP activations of our transformer model, with an MSENote that using an MSE loss avoids the challenges with polysemanticity that we discussed above in [Why Not Architectural Approaches?](https://transformer-circuits.pub/2023/monosemantic-features/#why-not-architectures) loss plus an L1 penalty to encourage sparsity.\nIn training the autoencoder, we found a couple of principles to be quite important. First, scale really matters. We found that training the autoencoder on more data made features subjectively “sharper” and more interpretable. In the end, we decided to use 8 billion training points for the autoencoder (see [Autoencoder Dataset](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-dataset)).\nSecond, we found that over the course of training some neurons cease to activate, even across a large number of datapoints. We found that “resampling” these dead neurons during training gave better results by allowing the model to represent more features for a given autoencoder hidden layer dimension. Our resampling procedure is detailed in [Neuron Resampling](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-resampling), but in brief we periodically check for neurons which have not fired in a significant number of steps and reset the encoder weights on the dead neurons to match data points that the autoencoder does not currently represent well.\nFor readers looking to apply this approach, we supply an appendix with [Advice for Training Sparse Autoencoders](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder).\n#### How can we tell if the autoencoder is working?\nUsually in machine learning we can quite easily tell if a method is working by looking at an easily-measured quantity like the test loss. We spent quite some time searching for an equivalent metric to guide our efforts here, and unfortunately have yet to find anything satisfactory.\nWe began by looking for [an information-based metric](https://transformer-circuits.pub/2023/may-update/index.html#simple-factorization), so that we could say in some sense that the best factorization is the one that minimizes the total information of the autoencoder and the data. Unfortunately, this total information did not generally correlate with subjective feature interpretability or activation sparsity. (Runs whose feature activations had an average L0 norm in the hundreds but low reconstruction error could have lower total information than those with smaller average L0 norm and higher reconstruction error.)\nThus we ended up using a combination of several additional metrics to guide our investigations:\n  1. Manual inspection: Do the features seem interpretable?\n  2. Feature density: we found that the number of “live” features and the percentage of tokens on which they fire to be an extremely useful guide. (See appendix for [details](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-density).)\n  3. Reconstruction loss: How well does the autoencoder reconstruct the MLP activations? Our goal is ultimately to explain the function of the MLP layer, so the MSE loss should be low.\n  4. Toy models: Having toy models where we know the ground truth and so can cleanly evaluate the autoencoder’s performance was crucial to our early progress.\n\n\nInterpreting or measuring some of these signals can be difficult, though. For instance, at various points we thought we saw features which at first didn’t make any sense, but with deeper inspection we [could understand](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs). Likewise, while we have identified some desiderata for the distribution of feature densities, there is much that we still do not understand and which prevents this from providing a clear signal of progress.\nWe think it would be very helpful if we could identify better metrics for dictionary learning solutions from sparse autoencoders trained on transformers.\n### [The (one-layer) model we're studying](https://transformer-circuits.pub/2023/monosemantic-features/#setup-transformer)\nWe chose to study a one-layer transformer model. We view this model as a testbed for dictionary learning, and in that role it brings three key advantages:\n  1. Because one-layer models are small they likely have fewer \"true features\" than larger models, meaning features learned by smaller dictionaries might cover all their \"true features\". Smaller dictionaries are cheaper to train, allowing for fast hyperparameter optimization and experimentation.\n  2. We can highly overtrain a one-layer transformer quite cheaply. We hypothesize that a very high number of training tokens may allow our model to learn cleaner representations in superposition.\n  3. We can easily analyze the effects features have on the logit outputs, because these are approximately linear in the feature activations.Note that this linear structure makes it even more likely that features should be linear. On the one hand, this means that the linear representation hypothesis is more likely to hold for this model. On the other hand, it potentially means that our results are less likely to generalize to multilayer models. Fortunately, others have studied multilayer transformers with sparse autoencoders and found interpretable linear features, which gives us more confidence that what we see in the one-layer model indeed generalizes . This provides helpful corroboration that the features we have found are not just telling us about the data distribution, and actually [reflect the functionality of the model](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-about-model).\n\n\nWe trained two one-layer transformers with the same hyperparameters and datasets, differing only in the random seed used for initialization. We then learned dictionaries of many different sizes on both transformers, using the same hyperparameters for each matched pair of dictionaries but training on the activations of different tokens for each transformer.\nWe refer to the main transformer we study in this paper as the “A” transformer. We primarily use the other transformer (“B”) to study [feature universality](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-universality), as we can e.g. compare features learned from the “A” and “B” transformers and see how similar they are.\n### [Notation for Features](https://transformer-circuits.pub/2023/monosemantic-features/#setup-notation)\nThroughout this draft, we'll use strings like \"A/1/2357\" to denote features. The first portion \"A\" or \"B\" denote which model the features come from. The second part (e.g. the \"1\" in \"A/1\") denotes the dictionary learning run. These vary in the number of learned factors and the L1 coefficient used. A table of all of our runs is available [here](https://transformer-circuits.pub/2023/monosemantic-features/vis/index.html). Notably, A/0…A/5 form a sequence with fixed L1 coefficients and increasing dictionary sizes. The final portion (e.g. the \"2357\" in \"A/1/2357\") corresponds to the specific feature in the run.\nSometimes, we want to denote neurons from the transformer rather than features learned by the sparse autoencoder. In this case, we use the notation \"A/neurons/32\".\n### [Interface for Exploring Features](https://transformer-circuits.pub/2023/monosemantic-features/#setup-interface)\nWe provide an interface for exploring all the features in all our dictionary learning runs. Links to the visualizations for each run can be found [here](https://transformer-circuits.pub/2023/monosemantic-features/vis/index.html). We suggest beginning with the [interface for A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html), which we discuss the most.\nThese interfaces provide extensive information on each feature. This includes examples of when they activate, what effect they have on the logits when they do, examples of how they affect the probability of tokens if the feature is ablated, and much more:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nOur interface also allows users to search through features:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nAdditionally, we provide a [second ](https://transformer-circuits.pub/2023/monosemantic-features/vis/index.html#example-texts)[interface](https://transformer-circuits.pub/2023/monosemantic-features/vis/index.html#example-texts) displaying all features active on a given dataset example. This is available for a set of example texts.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)   \n  \n  \n\n* * *\n  \n  \n\n## [Detailed Investigations of Individual Features](https://transformer-circuits.pub/2023/monosemantic-features/#feature-analysis)\nThe most important claim of our paper is that dictionary learning can extract features that are significantly more monosemantic than neurons. In this section, we give a detailed demonstration of this claim for a small number of features which activate in highly specific contexts.\nThe features we study respond to\n  * Text written in Arabic script (like \"الكتاب المختصر في حساب الجبر والمقابلة\")\n  * DNA sequences (like \"CCTGGTACTGTACGAACGAACGAACGTAGCCTTGG\")\n  * base64 strings (like the final characters in \"https://www.youtube.com/watch?v=dQw4w9WgXcQ\")\n  * Text written in Hebrew script (like \"בראשית ברא אלהים את השמים ואת הארץ\")\n\n\nFor each learned feature, we attempt to establish the following claims:\n  1. The learned feature activates with high specificity for the hypothesized context. (When the feature is on the context is usually present.)\n  2. The learned feature activates with high sensitivity for the hypothesized context. (When the context is present, the feature is usually on.)\n  3. The learned feature causes appropriate downstream behavior.\n  4. The learned feature does not correspond to any neuron.\n  5. The learned feature is universal – a similar feature is found by dictionary learning applied to a different model.\n\n\nTo demonstrate claims 1–3, we devise computational proxies for each context, numerical scores estimating the (log-)likelihood that a string (or token) is from the specific context. The contexts chosen above are easy to model based on the defined sets of unicode characters involved. We model DNA sequences as random strings of characters from `[ATCG]` and we model base64 strings as random sequences of characters from `[a-zA-Z0-9+/]`. For Arabic script and Hebrew features, we exploit the fact that each language is written in a script consisting of well-defined Unicode blocks. Each computational proxy is then an estimate of the log-likelihood ratio of a string under the hypothesis versus under the full empirical distribution of the dataset. The full description of how we estimate \\log(P(s|\\text{context}) / P(s)) for each feature hypothesis is given in the [appendix section on proxies](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-proxies).\nIn this section we primarily study the learned feature which is most active in each context. There are typically other features that also model that context, and we find that rare “gaps” in the sensitivity of a main feature are often explained by the activity of another. We discuss this phenomenon in detail in sections on [Activation Sensitivity](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-sensitivity) and [Feature Splitting](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-splitting).\nWe take pains to demonstrate the specificity of each feature, as we believe that to be more important for ruling out polysemanticity. Polysemanticity typically involves neurons activating for clearly unrelated concepts.  For example, [one neuron](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?ordering=index#feature-83) in our transformer model responds to a mix of academic citations, English dialogue, HTTP requests, and Korean text. In vision models, there is a classic example of a neuron which responds to cat faces and fronts of cars . If our proxy shows that a feature only activates in some relatively rare and specific context, that would exclude the typical form of polysemanticity.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe finally note that the features in this section are cherry-picked to be easier to analyze. Defining simple computational proxies for most features we find, such as text concerning [fantasy games](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html?search_target_mode=index_name&search_mode=case_insensitive#feature-129), would be difficult, and we analyze them in other ways in the [following section](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis).\n### [Arabic Script Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic)\nThe first feature we'll consider is an Arabic Script feature, [A/1/3450](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450). It activates in response to text in Arabic, Farsi, Urdu (and possibly other languages), which use the Arabic script. This feature is quite specific and relatively sensitive to Arabic script, and effectively invisible if we view the model in terms of individual neurons.\n#### [Activation Specificity](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-activations)\nOur first step is to show that this feature fires almost exclusively on text in Arabic script. We give each token an \"Arabic script\" score using an estimated likelihood ratio \\log(P(s|\\text{Arabic Script}) / P(s)), and break down the histogram of feature activations by that score. Arabic text is quite rare in our overall data distribution – just 0.13% of training tokens — but it makes up 81% of the tokens on which our feature is active. That percentage varies significantly by feature activity level, from 25% when the feature is barely active to 98% when the feature activation is above 5.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe also show dataset examples demonstrating different levels of feature activity. In interpreting them, it's important to note that Arabic Unicode characters are often split into multiple tokens. For example, the character `ث` ([U+062B](https://www.compart.com/en/unicode/U+062B)) is tokenized as `\\xd8` followed by `\\xab`.This kind of split tokenization is common for Unicode characters outside the Latin script and the most common characters from other Unicode blocks. When this happens, A/1/3450 typically only fires on the last token of the Unicode character.\nThe upper parts of the activation spectrum, above an activity of ~5, clearly respond with high specificity to Arabic script. What should we make of the lower portions? We have three hypotheses:\n  * The proxy is imperfect. It has false negatives due to common characters which are in other Unicode blocks (e.g., whitespace, punctuation), and also due to places where tokenization has created strange behavior.For example, in the figure above, there is a newline character ⏎ that our feature fires on but our proxy assigns a low score to because it is outside the Unicode block.\n  * The model may be imperfect (but still calibrated). If features activate proportional to their \"confidence\" that some property is present, then we should expect them to sometimes be wrong in their weak activations. Although it might seem silly to fail to recognize which language a piece of text is, this is a very weak one layer transformer model, and the fact that characters are often split into multiple tokens adds additional difficulty.\n  * The autoencoder may be imperfect: If the width of our autoencoder is less than the number of \"true features\" being used by the model, then unrecovered features may show up as low activations across many of our learned features.\n\n\nRegardless, large feature activations have larger impacts on model predictions,Why do we believe large activations have larger effects? In the case of a one-layer transformer like the one we consider in this paper, we can make a strong case for this: features have a linear effect on the logits (modulo rescaling by layer norm), and so a larger activation of the feature has a larger effect on the logits. In the case of larger models, this follows from a lot of conjectures and heuristic arguments (e.g. the abundance of linear pathways in the model and the idea of linear features at each layer), and must be true for sufficiently small activations by continuity, but doesn't have a watertight argument. so getting their interpretation right matters most. One useful tool for assessing the impact of false positives at low activation levels is the \"expected value plot\" of Cammarata et al. . We plot the distribution of feature activations weighted by activation level. Most of the magnitude of activation provided by this feature comes from dataset examples which are in Arabic script.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n#### [Activation Sensitivity](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-sensitivity)\nIn the Feature Activation Distribution above, it's clear that A/1/3450 is not sensitive to all tokens in Arabic script. In the random dataset examples, it fails to fire on five examples of the prefix \"ال\", transliterated as \"al-\", which is the equivalent of the definite article \"the\" in English. However, in exactly those places, another feature which is specific to Arabic script, [A/1/3134](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3134), fires. There are several additional features that fire on Arabic and related scripts (e.g. [A/1/1466](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1466), [A/1/3134](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3134), [A/1/3399](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3399)) which contribute to representing Arabic script. Another example deals with Unicode tokenization: when Arabic characters are split into multiple tokens, the feature we analyze here only activates at the final token comprising the character, while [A/1/3399](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3399) activates on the first token comprising the character. To see how these features collaborate, we provide an [alternative visualization](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1-ar.html) showing all the features active on a snippet of Arabic text. We consider such interactions more in the [Phenomenology](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology) section below.\nNevertheless, we find a Pearson correlation of 0.74 between the activity of our feature and the activity of the Arabic script proxy (thresholded at 0), over a dataset of 40 million tokens. Correlation provides a joint measure of sensitivity and specificity that takes magnitude into account, and 0.74 is a substantial correlation.\n#### [Feature Downstream Effects](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-effect)\nBecause the autoencoder is trained on model activations, the features it learns could in theory represent structure in the training data alone, without any relevance to the network’s function. We show instead that the learned features have interpretable causal effects on model outputs which make sense in light of the features’ activations. Note that these downstream effects are not inputs to the dictionary learning process, which only sees the activations of the MLP layer. If the resulting features also mediate important downstream behavioral effects then we can be confident that the feature is truly connected to the MLP’s functional role in the network and not just a property of the underlying data.\nWe begin with a linear approximation to the effect of each feature on the model logits. We compute the logit weight following the path expansion approach of , (also called “logit attribution”, see similar work e.g. ) multiplying each feature direction by the MLP output weights, an approximation of the layer norm operation combining a projection to remove the mean and a diagonal matrix approximating the scaling terms, and the unembedding matrix (d_iW_{down}\\pi L W_{unembed}). Because the softmax function is shift-invariant there is no absolute scale for the logit weights; we shift these so that the median logit weight for each feature is zero.\nEach feature, when active, makes some output tokens more likely and some output tokens less likely. We plot that distributionof logit weights.We exclude weights corresponding to extremely rare or never used vocabulary elements. These are perhaps similar to the \"anomalous tokens\" (e.g., \"SolidGoldMagikarp\") of Rumbelow & Watkins . There is a large primary mode at zero, and a second much smaller mode on the far right, the tokens whose likelihood most increases when our feature is on. This second mode appears to correspond to Arabic characters, and tokens which help represent Arabic script characters (especially `\\xd8` and `\\xd9`, which are often the first half of the UTF-8 encodings of Arabic Unicode characters in the [basic Arabic Unicode block](https://www.compart.com/en/unicode/block/U+0600)).\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThis suggests that activating this feature increases the probability the network predicts Arabic script tokens. There are in theory several ways the logit weights could overestimate the model's actual use of a feature:   \n1. It could be that these output weights are small enough that, when multiplied by activations, they don't have an appreciable effect on the model’s output.  \n2. The feature might only activate in situations where other features make these tokens extremely unlikely, such that the feature in fact has little effect.   \n3. It is possible that our approximation of linearizing the layer norm (see Framework ) is poor.  \nBased on the subsequent analysis, which confirms the logit weight effects, we do not believe these issues arise in practice.\nTo visualize these effects on actual data, we [causally ](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-ablations)[ablate](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-ablations)[ the feature](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-ablations). For a given dataset example, we run the context through the model until the MLP layer, decode the activations into features, then subtract off the activation of [A/1/3450](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450), artificially setting it to zero on the whole context, before applying the rest of the model. We visualize the effect of ablating the feature using underlines in the visualization; tokens whose predictions were helped by the feature (ablation decreased likelihood) are underlined in blue and tokens whose predictions were hurt by the feature (ablation increased likelihood) are underlined in red.\nIn the example on the right below we see that the [A/1/3450](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450) was active on every token in a short context (orange background). Ablating it hurt the predictions of all the tokens in Arabic script (purple underlines), but helped the prediction of the period `.` (orange underline). The rest of the figure displays contexts from two different ranges of feature activation levels. (The feature activation on the middle token of examples on the right (\"subsample interval 5\") is about half that of the middle token of examples on the left (\"subsample interval 0\")). We see that the feature was causally helping the model predictions on Arabic script through that full range, and the only tokens made less likely by the feature are punctuation shared with other scripts. The magnitudes of the impact are larger when the feature is more active.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe encourage interested readers to view the feature visualization for [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html) to review this and other effects.\nWe also validate that the feature's downstream effect is in line with our interpretation as an Arabic script feature by sampling from the model with the feature activity \"pinned\" at a high value. To do this, we start with a prefix `1,2,3,4,5,6,7,8,9,10` where the model has an expected continuation (keep in mind that this is a one layer model that is very weak!). We then instead set [A/1/3450](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450) to its maximum observed value and see how that changes the samples:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n#### [The feature is not a neuron](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-neuron)\nThis feature seems rather monosemantic, but some models have relatively monosemantic neurons, and we want to check that dictionary learning didn't merely hand us a particularly nice neuron.Of course, some features might genuinely be neuron aligned. But we'd like to know that at least some of the features dictionary learning discovers were not trivial. We first note that when we [search](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?search_mode=regex_case_sensitive&ordering=count&search_text=%5B%5Cu0600-%5Cu06FF%5D+) for neurons with Arabic script in their top dataset 20 examples, only one neuron turns up, with a single example in Arabic script. (Eighteen of the remaining examples are in English, and one is in Cyrillic.)\nWe then look at the coefficients of the feature in the neuron basis, and find that the three largest coefficients by magnitude are all negative (!) and there are a full 27 neurons whose coefficients are at least 0.1 in magnitude.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nIt is of course possible that these neurons engage in a delicate game of cancellation, resulting in one particular neuron's primary activations being sharpened. To check for this, we find the neuron whose activations are most correlated to the feature's activations over a set of ~40 million dataset examples. We also tried looking at the neurons which have the largest contribution to the dictionary vector for the feature. However, we found looking at the most correlated neuron to be more reliable – in earlier experiments, we found rare cases where the correlated method found a seemingly similar neuron, while the dictionary method did not. This may be because neurons can have different scales of activations. The most correlated neuron ([A/neurons/489](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html#feature-489)) responds to a mixture of different non-English languages.This makes sense as a superposition strategy: since languages are essentially mutually exclusive, they're natural to put in superposition with each other We can see this by visualizing the activation distribution of the neuron – the dataset examples are all other languages, with Arabic script present only as a small sliver.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nLogit weight analysis is also consistent with this neuron responding to a mixture of languages. For example, in the figure below many of the top logit weights appear to include Russian and Korean tokens. Careful readers will observe a thin red sliver corresponding to rare Arabic script tokens in the distribution. These Arabic script tokens have weight values that are very slightly positive leaning overall, but some are negative.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nFinally, scatter plots and correlations suggest the similarities between [A/1/3450](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450) and the neuron are non-zero, but quite minimal.By analogy, this also applies to [B/1/1334](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-1334). In particular, note how the y-axis of the logit weights scatter plot (corresponding to the feature) cleanly separates the Arabic script token logits, while the x-axis does not. More generally, note how the y-marginals both clearly exhibit specificity, but the x-marginals do not.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe conclude that the features we study do not trivially correspond to a single neuron. The Arabic script feature would be effectively invisible if we only analyzed the model in terms of neurons.\n#### [Universality](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-universality)\nWe will now ask whether A/1/3450 is a universal feature that forms in other models and can be consistently discovered by dictionary learning. This would indicate we are discovering something more general about how one-layer transformers learn representations of the dataset.\nWe search for a similar feature in [B/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html), a dictionary learning run on a transformer trained on the same dataset but with a different random seed. We search for the feature with the highest activation correlation \"Activation correlation\" is defined as the feature whose activations across 40,960,000 tokens has the highest Pearson correlation with those of [A/1/3450](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450). and find [B/1/1334](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-1334) (corr=0.91), which is strikingly similar:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThis feature clearly responds to Arabic script as well. If anything, it's nicer than our original feature – it's more specific in the 0–1 range. The logit weights tell a similar story:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe effects of ablating this feature are also consistent with this (see the visualization for [B/1/1334](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-1334)).\nTo more systematically analyze the similarities between A and B, we look at scatter plots comparing the activations or logit weights:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe activations are strongly correlated (Pearson correlation of 0.91), especially in the main Arabic mode.\nThe logit weights reveal a two-dimensional version of the bimodality we saw in the histogram for A/1, with logit weights for Arabic tokens clustering at the top right. The correlation is more modest than that of the activations because the distribution is dominated by a relatively uncorrelated mode in the center. We hypothesize this central mode corresponds to \"weight interference\" and that the shared outlier mode is the important observation – that is, the model may ideally prefer to have all those weights be zero, but due to superposition with other features and their weights, this isn't possible.\n### [DNA Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-dna)\nWe now consider a DNA feature, [A/1/2937](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2937). It activates in response to long uppercase strings consisting of `A`, `T`, `C`, and `G`, typically used to represent nucleotide sequences. We closely follow the analysis of the Arabic script feature above to show activation specificity and sensitivity for the feature, sensible downstream effects, a lack of neuron alignment, and universality between models. The main differences will be that (1) [A/1/2937](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2937) is the only feature devoted to modeling the DNA context, and (2) our proxy is less sensitive to DNA than our feature is, missing strings containing punctuation, spaces, and missing bases.\n#### [Activation Specificity and Sensitivity](https://transformer-circuits.pub/2023/monosemantic-features/#feature-dna-activations)\nWe begin with the computational proxy for \"is a DNA sequence\", \\log(P(s|\\text{DNA}) / P(s)). Because there are some vocabulary tokens, such as `CAT`, which could occur in DNA but also occur in other contexts, we always look at groups of at least two tokens when evaluating the proxy. The log-probabilities turn out to be quite bimodal, so we binarize the proxy (based on its sign). This binarized proxy then has a Pearson correlation of 0.8 with the feature activations.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWhile the feature appears to be quite monosemantic in the feature's higher registers (all 10 random dataset examples above activation of 6.0 are DNA sequences), there is significant blue indicating the DNA proxy not firing in the lower registers. Below we show a grid of random examples at four activation levels (including feature off) where the proxy does and doesn't fire.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe note that in all but two cases where the feature and proxy disagree, the feature does indicate a DNA sequence, just one outside our proxy's strict `ATCG` vocabulary. For example, the space present in the triplets ` TGG AGT` makes the proxy fail to fire. The feature also fires productively on `'-`in the string `5'-TCT`, because what follows that prefix should be DNA, even though the prefix is not itself DNA. (The causal ablation reveals that turning off the DNA feature hurts the prediction of the strings that follow.) We thus believe that [A/1/2937](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2937) is quite sensitive and specific for DNA. The case where the proxy fires and the feature does not is indeed a DNA sequence, though the feature begins firing on the very next token. We observe that the DNA feature may not fire on the first few tokens of a DNA sequence, [but by the end of a long DNA sequence](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1-dna.html), it is the only feature active.\n#### [Feature Downstream Effect](https://transformer-circuits.pub/2023/monosemantic-features/#feature-dna-effect)\nThe downstream effect of the DNA feature being active, as measured by logit weights, make sense, with all the top tokens being combinations of nucleotides like `AGT` and `GCC`.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n#### [The Feature Is Not A Neuron](https://transformer-circuits.pub/2023/monosemantic-features/#feature-dna-neuron)\nThe most similar neuron to [A/1/2937](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2937), as measured by activation correlation, is [A/neurons/67](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html#feature-67). DNA contexts form a tiny sliver of that neuron's activating examples. The neuron whose coefficient is the highest in our feature's vector, [A/neurons/227](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html#feature-227), also has no DNA sequences in its top activating examples.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n#### [Universality](https://transformer-circuits.pub/2023/monosemantic-features/#feature-dna-universality)\n[A/1/2937](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2937) has a correlated feature (corr=0.92) in run B/1, [B/1/3680](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-3680). Their top logit weights agree, and are DNA tokens (e.g. `AGT`) forming a separate mode (circled on the right) than the bulk of the logit weights for both.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n### [Base64 Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-base64)\nWe now consider a [base64](https://en.wikipedia.org/wiki/Base64) feature, [A/1/2357](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357). We're particularly excited by this feature because we discovered a base64 neuron in our SoLU paper , suggesting that base64 might be quite universal – even across models trained on somewhat different datasets and with different architectures. We model base64 strings as random sequences of characters from `[a-zA-Z0-9+/]`. The activation distribution colored by the corresponding computational proxy, together with the random dataset examples from each activation level, shows that this feature is quite specific to base64.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThis is not the only feature active in base64 contexts, and in a section [below](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs-2) we discuss the two others, one of which fires on single digits in base64 contexts (like the 2, 4, 7, and 9 on which A/1/2357 doesn't activate in the figure above), exploiting a property of the BPE tokenizer to make a better prediction.\nTurning to the logit weights, they have a second mode consisting of highly base64-specific tokens. The main mode seems to primarily be interference, but the right side is skewed towards base64-neutral or slightly base64-leaning tokens. (If we look at the conditional below, we see a more continuous transition to base64-specific tokens.)\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThere is a more continuous transition between non-base64 and base64 tokens than we saw in the Arabic script example. This difference likely arises because whether a token occurs more in Arabic script than in other text is a relatively binary distinction, whereas whether a token occurs more in base64 or other text varies more continuously. For instance, `fr` is both a common abbreviation for the French language and also a base64 token, so it makes sense for the model to be cautious in up-weighting `fr` because it might already have a higher prior due to use in French. Indeed, any token consisting of letters from the English alphabet will have some nontrivial probability of appearing in base64 strings.\nThe Pearson correlation between the computational proxy and the activity of [A/1/2357](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357) is just 0.38. We believe that is mostly because the proxy is too broad. For example hexadecimal strings (those made of `[0-9A-F]`) activate the proxy, as they are quite different from the overall data distribution, but are actually predicted by a feature of their own, [A/1/3817](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3817).\n#### [Universality](https://transformer-circuits.pub/2023/monosemantic-features/#feature-base64-universality)\n[A/1/2357](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357) has a correlated feature (corr=0.85) in run B/1, [B/1/2165](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-2165). It also has high activation specificity for base64 strings:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nLike [A/1/2357](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357), [B/1/2165](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-2165)'s logit weights have a second mode corresponding to base64 tokens:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nCorrelations and scatter plots are also consistent with them being very similar features:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nNote that we expect the overlap between the interference and base64 token logit weights to be from the aforementioned usage of base64 tokens across many other contexts.\n#### [The Feature Is Not A Neuron](https://transformer-circuits.pub/2023/monosemantic-features/#feature-base64-neuron)\nLooking at the neuron in model A that most correlates with this feature: [A/neurons/470](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html#feature-470) (corr=0.18), we find that while it does notably respond to base64 strings, it also activates for lots of other things, including code, HTML labels, parts of URLs, etc.:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe logit weights suggest it somewhat increases base64 tokens, but is much more focused on upweighting other tokens, e.g. filename endings.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe activation and logit correlations are consistent with this neuron helping represent the same feature, but largely doing other things.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n### [Hebrew Feature](https://transformer-circuits.pub/2023/monosemantic-features/#feature-hebrew)\nAnother interesting example is the Hebrew feature [A/1/416](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-416). Like the Arabic feature, it's easy to computationally identify Hebrew text based on Unicode blocks.\n[A/1/416](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-416) has high activation specificity in the upper spectrum. It does weakly activate for other things (especially other languages with Unicode scripts). There is also some blue in strong activations; this appears to significantly be on \"common characters\", such as whitespace or punctuation, which are from other unicode blocks (see more discussion of similar issues in the [Arabic feature section](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-activations)).\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nIts logit weights have a notable second mode, corresponding to Hebrew characters and relevant incomplete Unicode characters. Note that `\\xd7` is the first token in the UTF-8 encoding of most characters in the [basic Hebrew Unicode block](https://www.compart.com/en/unicode/block/U+0590).\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe Pearson correlation of the Hebrew script proxy with [A/1/416](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-416) is 0.55. Some of the failure of sensitivity may be due to a complementary feature[ A/1/1016](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1016) that fires on `\\xd7` and predicts the bytes that complete Hebrew characters' codepoints.\n#### [The Feature Is Not A Neuron](https://transformer-circuits.pub/2023/monosemantic-features/#feature-hebrew-neuron)\nThere doesn't appear to be a similar neuron. The most correlated neuron in model A is [A/neurons/489](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?ordering=index#feature-489) (corr=0.1), which has low activation and logit specificity. Consider the following activation and logit correlation plots:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nTo cross-validate this, we also [searched](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?search_mode=regex_case_sensitive&ordering=count&search_text=%5B%5Cu0590-%5Cu05FF%5D) for any neuron where the main Hebrew Unicode block appeared in the top dataset examples. We found none.\n#### [Universality](https://transformer-circuits.pub/2023/monosemantic-features/#feature-hebrew-universality)\n[A/1/416](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-416) has a correlated feature in the B/1 run, [B/1/1901](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-1901) (corr=0.92) that has significant activation specificity:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nLogit weights have a second mode, as before:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nActivation and logit weight correlations are again consistent:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)   \n  \n  \n\n* * *\n  \n  \n\n## [Global Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis)\nIf the previous section has persuaded you that at least some of the features are genuinely interpretable and reflect the underlying model mechanics, it's natural to wonder how broadly this holds outside of those cherry-picked features. The primary focus of this section will be to answer the question, \"how interpretable are the rest of the features?\" We show that both humans and large language models find our features to be significantly more interpretable than neurons, and quite interpretable in absolute terms.\nThere are a number of other questions one might also ask. To what extent is our dictionary learning method discovering all the features necessary to understand the MLP layer? Holistically, how much of the MLP layer's mechanics have been made interpretable? We are not yet able to fully answer these questions to our satisfaction, but will provide some preliminary speculation towards the end of this section.\nWe note that of the 4,096 learned features in the A/1 autoencoder, 168 of them are \"dead\" (active on none of the 100 million dataset examples) and 292 of them are \"ultralow density\", active on less than 1 in a million dataset examples and exhibiting [other atypical properties](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-density). We exclude both these groups of features from further analyses.\n### [How Interpretable is the Typical Feature?](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp)\nIn this section, we use three different methods to analyze how interpretable the typical feature is, and how that compares to neurons: human analysis, and two forms of automated interpretability. All three approaches find that features are much more interpretable than neurons.\n#### [Manual Human Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-manual)\nAt present, we do not have any metric we trust more than human judgment of interpretability. Thus, we had a blinded annotator (one of the authors, Adam Jermyn) score features and neurons based on how interpretable they are. The scoring rubric can be found in the [appendix](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-rubric) and accounts for confidence in an explanation, consistency of the activations with that explanation, consistency of the logit output weights with that explanation, and specificity.\nIn doing this evaluation, we wanted to avoid a weakness we perceived in our prior work (e.g., ) of focusing evaluation predominantly on maximal dataset examples, and paying less attention to the rest of the activation spectrum. Many polysemantic neurons appear monosemantic if you only look at top dataset examples, but are revealed to be polysemantic if you look at lower parts of the activation spectrum. To avoid this, we draw samples uniformly across the spectrum of feature activations,In order to sample uniformly across the spectrum of feature activations, we divide the activation spectrum into 11 \"activation intervals\" evenly spaced between 0 activation and the maximum activation. We sample uniformly from these intervals. and score each interval separately in light of the overall hypothesis suggested by the feature.\nUnfortunately, this approach is labor intensive and so the number of scored samples is small. In total, 412 feature activation intervals were scored across 162 features and neurons.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe see that features are substantially more interpretable than neurons. Very subjectively, we found features to be quite interpretable if their rubric value was above 8. The median neuron scored 0 on our rubric, indicating that our annotator could not even form a hypothesis of what the neuron could represent! Whereas the median feature interval scored a 12, indicating that the annotator had a confident, specific, consistent hypothesis that made sense in terms of the logit output weights. \n#### [Automated Interpretability – Activations](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-auto-acts)\nTo analyze features at a larger scale, we turned to automated interpretability . Following the approach of Bills et al. , we have a large language model, Anthropic’s Claude, generate explanations of features using examples of tokens where they activate. Next, we have the model use that explanation to predict new activations on previously unseen tokens.It's worth explicitly stating that our automated interpretability setup was designed to ensure that there's no leak of information about activation patterns, except for the explanation. For example, when predicting new activations, the model cannot see any true activations of that feature.\nLike with the human analysis, we used samples across the full range of activation intervals to evaluate monosemanticity.This is distinct from the evaluation strategy of Bills et al., who calculated their correlation of predicted and true activations on a mixture of maximal dataset examples and random samples. For sparse features, which don't fire on most random samples, this effectively tests the model's ability to distinguish a feature's large activations from zero. Concretely, for each feature, we computed the Spearman correlation coefficient between the predicted activation and the true activations for 60 dataset examples made up of nine tokens each, resulting in 540 predictions per feature. While only using completely random sequences would be the most principled approach to scoring, half of the examples are from across the feature intervals to get a more accurate correlation. See [the appendix](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-automated) for additional information including the use of importance scoring to precisely counter-weight the bias of providing tokens the feature fires for.\nIn agreement with the human analysis, Claude is able to explain and predict activations for features significantly better than for neurons.In instances where Claude predicts a constant score, most often all 0s, a correlation can't be computed and we assign a score of zero which explains the uptick there.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n#### [Automated Interpretability – Logit Weights](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-auto-logits)\nIn our earlier analysis of individual features, we found that looking at the logits is a powerful tool for cross-validating the interpretability of features. We can take this approach in automated interpretability as well. Using the explanations of features generated in the previous analysis, we ask a language model to predict if a previously unseen logit token is something the feature should predict as likely to come next. This is then scored against a 50/50 mix of top positive logit tokens and random other logit tokens. Randomly guessing would give a 50% accuracy, but the model instead achieves a 74% average across features, compared to a 58% average across neurons. Failures here refer to instances where Claude failed to reply in the correct format for scoring.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n#### [Activation Interval Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-intervals)\nIn addition to studying features as a whole, in our manual analysis we can zoom in on portions of the feature activation spectrum using the feature intervals. As before, a feature interval is the set of examples with activations closest to a specific evenly-spaced fraction of the max activation. So, rather than asking if a feature seems interpretable, we ask whether a range of activations is consistent with the overall hypothesis suggested by the full spectrum of the feature’s activation. This allows us to ask how interpretability changes with feature activation strength.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nHigher-activating feature intervals were more consistent with our interpretations than lower-activating ones. In particular:\n  * Many features show consistent activations across the entire activation spectrum.\n  * Some features show consistent activations across the top ~60% of the activation spectrum, and then quickly become less interpretable as we look to smaller and smaller activations.\n\n\nIt is possible that this is a sign that our features are not quite right. For instance, if one of our features is at a slight angle to the feature we’d really like to have learned, that can show up as inconsistent behavior in the lower activation intervals.\n#### [Caveats](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-interp-caveats)\nOur manual and automated interpretability experiments have a few caveats:\n  * Feature activations are skewed towards the lower intervals. Most feature activations are quite small, and so fall in the lower (and less-interpretable) feature intervals. That said, as we found in our detailed analysis, most of the effect of a featurewith respect to the effect it has on the model outputs see e.g. the activation expected value plot in [Arabic Feature's Activations Specificity Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-activations). is due to its higher activations, which are quite interpretable.\n  * The evaluated features are sampled uniformly from among the set of all features, and interpretability might be correlated with importance. One could imagine evaluating the features with the largest magnitude of activations, or the largest effect when ablated. Our sense is that these would be more interpretable than the randomly sampled features we considered. One could imagine a situation where the most important features in some sense were systematically less interpretable, though inspection of the [most active features](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html?ordering=max_dense) suggests the opposite.\n\n\nBased on our inspection of many features in the visualization, we believe these caveats do not affect the experimental results. We encourage interested readers to open the visualization for [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html) and the [corresponding neurons](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html). You can sort by ‘random’ to get an unbiased sample and do your own version of the above experiment, or sort features by importance metrics such as max activation and max density to evaluate the final caveat above.\n### [How much of the model does our interpretation explain?](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-how-much)\nWe now turn to the question we're least able to answer – to what extent do these seemingly interpretable features represent the \"full story\" of the MLP? One could imagine posing this question in a variety of ways. What fraction of the MLP loss contribution have we made interpretable? How much model behavior can we understand? If there really are some discrete set of \"true features\", what fraction have we discovered?\nOne way to partly get at this question is to ask how much of the loss is explained by our features. For [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html), the run we've focused most on in this paper, 79% of the log-likelihood loss reduction provided by the MLP layer is recovered by our features. That is, the additional loss incurred by replacing the MLP activations with the autoencoder's output is just 21% of the loss that would be incurred by zero ablating the MLP. This loss penalty can be reduced by using more features, or using a lower L1 coefficient. As an extreme example, [A/](https://transformer-circuits.pub/2023/monosemantic-features/vis/a5.html)[5](https://transformer-circuits.pub/2023/monosemantic-features/vis/a5.html) (`n_learned_sparse=131,072`, `l1_coefficient=0.004`) recovers 94.5% of log-likelihood loss.\nThese numbers should be taken with a significant grain of salt. The biggest issue is that framing this question in terms of fraction of loss may be misleading – we expect there to be a long-tail of features such that as the fraction of loss explained increases, more and more features are needed to explain the residual. Another issue is that we don't believe our features are completely monosemantic (some polysemanticity may be hiding in low activations), nor are all of them necessarily cleanly interpretable. With all of that said, our [earlier analyses](https://transformer-circuits.pub/2023/monosemantic-features/#feature-analysis) of individual features (e.g. the Arabic feature, base64 feature, etc.) do show that specific interpretable features are used by the model in interpretable ways – ablating them decreases probabilities in the appropriate way, and artificially activating them causes a corresponding behavior. This seems to confirm that the 79% of loss recovered is measuring something real, despite these caveats.\nIn principle, one could use automated interpretability to produce a better measure here: replacing activations with those predicted from explanations. (We believe others in the community have recently been considering this!) The naive versions of this would be quite computationally expensive,Naively, simulating a layer with 100k features would be 100,000 times more expensive than sampling a large language model such as Claude 2 or GPT-4 (we'd need to sample the explaining large model once for every feature at every token). At current prices, this would suggest simulating 100k features on a single 4096 token context would cost $12,500–$25,000, and one would presumably need to evaluate over many contexts. although there may be approximations. More generally, there is a much broader space of possibilities here. Perhaps there's a principled way to do this analysis in terms of single features – there are significant conceptual issues, but one might be able to formalize earlier notions of \"feature importance\" as an independent loss contribution a feature makes, and then analyze how much of that specific feature can be recovered.\nOverall, we view the problem of measuring the degree to which a feature-based interpretation explains a model to be an important open question, where significant work is necessary on both defining metrics and finding efficient ways to compute them.\n### [Do features tell us about the model or the data?](https://transformer-circuits.pub/2023/monosemantic-features/#global-analysis-about-model)\nA model's activations reflect two things: the distribution of the dataset and the way that distribution is transformed by the model. Dictionary learning on activations thus mixes data and model properties, and intriguing properties of learned features may be attributed to either or both sources. Correlations in the data can persist after application of the first part of the model (up to the MLP), and it is in theory possible that the intriguing features we see are merely artifacts of dataset correlations projected into a different space. However, the use of those features by the second half of the model (MLP downprojection and unembedding) are not an input to dictionary learning, so the interpretability of the downstream effects of those features must be a property of the model.\nTo assess the effect of dataset correlations on the interpretability of feature activations, we run dictionary learning on a version of our one-layer model with random weights.We generate a model with random weights by randomly shuffling the entries of each weight matrix of the trained transformer used in Run A. This guarantees that the distributions of individual weights match, and differences are due to structure. The resulting features are [here](https://transformer-circuits.pub/2023/monosemantic-features/vis/random1.html), and contain many single-token features (such as \"span\", \"file\", \".\", and \"nature\") and some other features firing on seemingly arbitrary subsets of different broadly recognizable contexts (such as [LaTeX](https://transformer-circuits.pub/2023/monosemantic-features/vis/random1.html#feature-1143) or [code](https://transformer-circuits.pub/2023/monosemantic-features/vis/random1.html#feature-3050)). However, we are unable to construct interpretations for the non-single-token features that make much sense and invite the reader to examine feature visualizations from the model with randomized weights to confirm this for themselves. We conclude that the learning process for the model creates a richer structure in its activations than the distribution of tokens in the dataset alone.\nTo assess the interpretability of the downstream feature effects, we again use the three main approaches of the previous section:\n  1. Logit weight inspection. The logit weights represent the effect of each feature on the logits, and, as demonstrated in our earlier [investigations of individual features](https://transformer-circuits.pub/2023/monosemantic-features/#feature-analysis), they are consistent with the feature activations for the base64, Arabic, and Hebrew features. The reader is invited to inspect logit weights for all features in the visualization.\n  2. Feature ablation. We set the value of a feature to zero throughout a context, and record how the loss on each token changes. Ablations are available for all features in the visualization.\n  3. Pinned feature sampling. We artificially pinned the value of a feature to a fixed high number and then sample from the model. We find that the generated text matches the interpretation of the feature.\n\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe empirical consistency of feature activations with their downstream effects across all these metrics provides evidence that the features found are being used by the model.\n  \n  \n  \n\n* * *\n  \n  \n\n## [Phenomenology](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology)\nUltimately, the goal of our work is to understand neural networks. Decomposition of models into features is simply a means to this end, and one might very reasonably wonder if it's genuinely advancing our overall goal. So, in this section, we'll turn attention to the lessons these features can teach us about neural networks. (We've taken to calling this work of leveraging our theoretical understanding to reason about model properties phenomenology by analogy to [phenomenology in physics](https://en.wikipedia.org/wiki/Phenomenology_\\(physics\\)), and the [2019 ICML workshop](https://deep-phenomena.org/) on phenomena in deep learning.)\nOne way to do this would be to give a detailed discussion of the features we've found (similar to ), but we believe the best way to get a sense of the features we discovered is simply to [browse the interface](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html) for exploring features which we've published alongside this paper. The features we find vary enormously, and no concise summary will capture their breadth. Instead, we will largely focus on more abstract properties and patterns that we notice. These abstract properties will be able to inform – although by no means answer – questions like \"Are these the real features?\" and \"What is actually going on in a one-layer model?\"\nWe begin by discussing some basic motifs and observations about features. We'll then discuss how the features we relate compare to features in other dictionary learning runs and in other models. This will suggest that features are universal and that dictionary learning can be understood as a process of feature splitting that reflects something deep about the geometry of superposition. Finally, we'll explore how features connect together into \"finite state automata\" as systems that implement more complex behaviors.\n### [Feature Motifs](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-motifs)\nWhat kinds of features do we find in our model?\nOne strong theme is the prevalence of context features (e.g. DNA, base64) and token-in-context features (e.g. `the` in mathematics – [A/0/341](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-341), `<` in HTML – [A/0/20](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-20)).From a purely theoretical lens, attention heads can largely implement \"three point functions\" (with two inputs, and an output). MLP layers are well positioned to instead implement N-token conjunctions, perhaps the most extreme of which are context features or token-in-context features. Thus, it is perhaps natural that we see many of these. These have been observed in prior work (context features e.g. ; token-in-context features e.g. ; preceding observations ), but the sheer volume of token-in-context features has been striking to us. For example, in [A/4](https://transformer-circuits.pub/2023/monosemantic-features/vis/a4.html), there are over a hundred features which primarily respond to the token \"the\" in different contexts.Token-in-context features may offer a significant opportunity for simplifying analysis of the model – as Elhage et al. note, it may be possible to understand these features as two-dimensional family of features parameterized by a context and a token. Often these features are connected by feature splitting (discussed in the next section), presenting as pure context features or token features in dictionaries with few learned features, but then splitting into token-in-context features as more features are learned.\nAnother interesting pattern is the implementation of what seem to be \"trigram\" features, such as a feature that predicts the `19` in `COVID-19` ([A/2/12310](https://transformer-circuits.pub/2023/monosemantic-features/vis/a2.html#feature-9143)). Such features could in principle be implemented with attention alone, but in practice the model uses the MLP layer as well. We also see features which seem to respond to specific, longer sequences of tokens. These are particularly striking because they may implement \"memorization\" like behavior – we'll discuss this more later.\nFinally, it's worth noting that all the features we find in a one-layer model can be interpreted as \"action features\" in addition to their role as \"input features\". For example, a base64 feature can be understood both as activating in response to base64 strings, and also as acting to increase the probability of base64 strings. The \"action\" view can clarify some of the token-in-context features: the feature [A/0/341](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-341) predicts noun phrases in mathematical text, upweighting nouns like `denominator` and adjectives like `latter`. Consequently, while it activates most strongly on `the`, it also activates on adjectives like `special` and `this` which are also followed by noun phrases. This dual interpretation of features can be explored by browsing our interface. Several papers have previously explored interpreting neurons as actions (e.g. ), and one-layer models are particularly suited to this, since it's a particularly principled way to understand the last MLP layer, and the only MLP in a one-layer model is the last layer.\n### [Feature Splitting](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-splitting)\nOne striking thing about the features we’ve found is that they appear in clusters. For instance, we observed above multiple base64 features, multiple Arabic script features, and so on. We see more of these features as we increase the total number of learned sparse features, a phenomenon we refer to as feature splitting. As we go from 512 features in A/0 to 4,096 features in A/1 and to 16,384 features in A/2, the number of features specific to base64 contexts goes from 1 to 3 to many more.\nTo understand how the geometry of the dictionary elements correspond to these qualitative clusters, we do a 2-D UMAP on the combined set of feature directions from A/0, A/1, and A/2.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe see clusters corresponding to the base64 and Arabic script features, together with many other tight clusters from specific contexts and a variety of other interesting geometric structures for other features. This confirms that the qualitative clusters are reflected in the geometry of the dictionary: similar features have small angles between their dictionary vectors.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe conjecture that there is some idealized set of features that dictionary learning would return if we provided it with an unlimited dictionary size. Often, these \"true features\" are clustered into sets of similar features, which the model puts in very tight superposition. Because the number of features is restricted, dictionary learning instead returns features which cover approximately the same territory as the idealized features, at the cost of being somewhat less specific.\nIn this picture, the reason the dictionary vectors of conceptually similar features are similar is that they are likely to produce similar behaviors in the model, and so should be responsible for similar effects in the neuron activations. For instance, it would be natural for a feature that fires on periods to predict tokens with a leading space followed by a capital letter. If there are multiple features that fire on periods, perhaps on periods in somewhat different contexts, these might all predict tokens with a leading space, and those predictions might well involve producing similar neuron activations. The combination of features being highly correlated and having similar \"output actions\", causes real models to have both denser and more structured superposition than what we observed in our previous toy models work .Toy Models considered correlated features (see in particular [organization of correlated features](https://transformer-circuits.pub/2022/toy_model/index.html#geometry-organization) and [collapsing of correlated features](https://transformer-circuits.pub/2022/toy_model/index.html#geometry-collapsing)), but only features which were correlated in whether they were active (and not their value if active), and had nothing analogous to the similar \"output actions\" described here. Nonetheless, Toy Models' experiments may be a useful intuition pump, especially in noticing the distinction between similar features in superposition vs features collapsing into a single broader feature.\nIf this picture is true, it would be important for a number of reasons. It suggests that determining the \"correct number of features\" for dictionary learning is less important than it might initially seem. It also suggests that dictionary learning with fewer features can provide a \"summary\" of model features, which might be very important in studying large models. Additionally, it would explain some of the stranger features we observe in the process of dictionary learning, suggesting that these are either \"collapsed\" features which would make sense if split further (see [\"Bug\" 1: Single Token Features](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs-1)), or else highly-specific \"split\" features which do in fact make sense if analyzed closely (see [\"Bug\" 2: Multiple Features for a Single Context](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs-2)). Finally, it suggests that our basic theory of superposition in toy models is missing an important dimension of the problem by not adequately studying highly correlated and \"action sharing\" features. \n#### Example: Mathematics and Physics Features\nIn this example, our coarsest run (with 512 learned sparse features) has three features describing tokens in different technical settings. Using the masked cosine similarityFor each pair of features, we compute the cosine similarity between their activations on the subset of tokens for which one of them fires. We repeat this, restricting to the subset on which the other fires, and take the greater of the two. We draw a connection between the features if this measure exceeds ~0.4. This threshold was chosen to balance producing a small enough graph to visualize while also showing some of the richness of feature splitting. We omit the ultralow density features from this analysis. between feature activations, we are able to identify how these features refine and split in runs with more learned sparse features.\nWhat we see is that the finer runs reveal more fine-grained distinctions between e.g. concepts in technical writing, and distinguish between the articles `the` and `a`, which are followed by slightly different sets of noun phrases. We also see that the structure of this refinement is more complex than a tree: rather, the features we find at one level may both split and merge to form refined features at the next. In general though, we see that runs with more learned sparse features tend to be more specific than those with fewer.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nIt's worth noting that these more precise features reflect differences in model predictions as well as activations. The general `the` in mathematical prose feature ([A/0/341](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-341)) has highly generic mathematical tokens for its top positive logits (e.g. supporting `the denominator`, `the remainder`, `the theorem`), whereas the more finely split machine learning version ([A/2/15021](https://transformer-circuits.pub/2023/monosemantic-features/vis/a2.html#feature-15021)) has much more specific topical predictions (e.g. `the dataset`, `the classifier`). Likewise, our abstract algebra and topology feature ([A/2/4878](https://transformer-circuits.pub/2023/monosemantic-features/vis/a2.html#feature-4878)) supports `the quotient` and `the subgroup`, and the gravitation and field theory feature ([A/2/2609](https://transformer-circuits.pub/2023/monosemantic-features/vis/a2.html#feature-2609)) supports `the gauge`, `the Lagrangian`, and `the spacetime`. \n#### [Features which seemed like Bugs](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs)\n[\"Bug\" 1: Single-Token Features](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs-1)\nWhen we limit dictionary learning to use very few learned sparse features, the features that emerge sometimes look quite strange. In particular, there are a large number of high-activation magnitude features which each only fire on a single token, and which seem to fire on every instance of that token. Such features are strange because the model could achieve the same effect entirely by learning different bigram statistics, and so should have no reason to devote MLP capacity to these. Similar features were also recently observed in a report by Smith .\nWe believe that feature splitting explains this phenomenon: the model hasn’t learned a single feature firing on the letter `P`,The features fire on the token \" P\" of words where it appears as the first token, such as `[ P][attern]`. for instance. Rather, it’s learned many features which fire on `P` in different contextsIn this instance we found the refined `P` features by manual inspection rather than by cosine similarity., with correspondingly different effects on the neuron activations and output logits (see below). At a sufficiently coarse level dictionary learning cannot tell the difference between these, but when we allow it to use more learned sparse features, the features split and refine into a zoo of different `P` features that fire in different contexts.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n[\"Bug\" 2: Multiple Features for a Single Context](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs-2)\nWe also observed the converse, where multiple features seemed to cover roughly the same concept or context. For example, there were three features in [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html) which fired on (subsets of) base64 strings, and predicted plausible base64 tokens like `zf`, `mF`, and `Gp`. One of these features was discussed in detail [earlier](https://transformer-circuits.pub/2023/monosemantic-features/#feature-base64), where we showed it fired for base64 strings. But we also observed that it didn't fire for all base64 strings – why? And what are the other two features doing? Why are there three?\nIn [A/0](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html) (with 512 features), the story is simple. There is only one base64-related feature, [A/0/45](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-45), which seems to activate on all tokens of base64-encoded strings. But in [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html), that feature splits into three different features whose activations seem to jointly cover those of [A/0/45](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-45):\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nTwo of these features seem relatively straightforward. [A/1/2357](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357) seems to fire preferentially on letters in base64, while [A/1/2364](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2364) seems to fire preferentially on digits.\nComparing the logit weights of these features reveals that they predict largely the same sets of tokens, with one significant difference: the feature that fires on digits has much lower logit weights for predicting digits. Put another way, if the present token is made of digits, the model will predict that the next token is a non-digit base64 token.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe believe this is likely an artifact of tokenization! If a single digit were followed by another digit, they would have been tokenized together as a single token; `[Bq][8][9][mp]` would never occur, as it would be tokenized instead as `[Bq][89][mp]`. Thus even in a random base64 string, the fact that the current token is a single digit gives information about the next token.\nBut what about the third feature, [A/1/1544](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1544)? At first glance, there isn't an obvious rule for when it fires. But if we look more closely, we notice that it seems to respond to base64 strings which encode ASCII text.Our initial clue that [A/1/1544](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1544) might fire on base64 strings encoding ASCII text was the token `ICAgICAg` which this feature particularly responds to, and corresponds to six spaces in a row. If we look at the top dataset examples for each feature, we find that examples for [A/1/1544](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1544) contain substrings which decode as ASCII, while none of the top activating examples for [A/1/2357](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357) or [A/1/2364](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2364) do:Determining when a dataset example encodes ASCII text is somewhat subtle because base64 can only be decoded to ASCII in groups of four characters, since four base64 characters [encode triples](https://en.wikipedia.org/wiki/Base64#Examples) of ASCII characters. Thus, we select substrings which – when decoded with the python base64 library – contain the maximal number of printable ASCII characters.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThis pattern of investigation, where one looks at coarser sets of features to understand categories of model behavior, and then at more refined sets of features to investigate the subtleties of that behavior, may prove well adapted to larger models where the feature set is expected to be quite large.\nIt's also worth noting how dictionary learning features were able to surprise us here. Many approaches to interpretability are top-down, and look for things we expect. But who would have known that models not only have a base64 feature, but that they distinguish between distinct kinds of base64 strings? This reminds us of cases like high-low frequency detectors or multimodal neurons where surprising and unexpected features were discovered in vision models.\n### [Universality](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-universality)\nOne of the biggest \"meta questions\" about features is whether they're [universal](https://distill.pub/2020/circuits/zoom-in/#claim-3) – do the same features form across different models? This question is generally important because it bears on whether the hard-earned lessons from studying one model will generalize to others. But it's especially important in the context of attempting to extract features from superposition because universality could provide significant evidence that the features we're extracting are \"real\", or at least reproducible.In what sense does universality suggest features are \"real\"? One basic observation is that they suggest the features we're finding are not just artifacts of the dictionary learning process – or at least that if they come from the dictionary learning process, it's in some consistent way. But there are also several deeper ways in which it's suggestive. It means that, whatever the source of the features, we can talk about features as replicable, reliable, recurring units of analysis. It's also just a surprising observation that one would expect if a strong version of the features in superposition hypothesis was true and models were literally representing some finite, discrete set of features in superposition.\nEarlier, we saw that all the features we performed detailed analyses of (e.g. the Arabic feature, or base64 feature) were universal between two one-layer models. But is this true for typical features in our model? And how broadly is it true – do we only observe the same feature if we train models of the same architectures on the same dataset, or do these features also occur in more divergent models? This section will seek to address these two questions. The first subsection will quantitatively analyze how widespread universality is between the two one-layer models we studied, while the second will compare the features we find to others reported in the literature in search of a stronger form of universality.\nWe observe substantial universality of both types.In fact, we found some features so universal that we began to take it for granted as a basic tool in our workflow of evaluating dictionary learning runs. For example, the base64 feature – which we previously observed in SoLU models – was so consistently universal that its presence was a useful debugging heuristic. At a high-level, this makes sense: if a feature is useful to one model in representing the dataset, it's likely useful to others, and if two models represent the same feature then a good dictionary learning algorithm should find it.\n#### Comparing features between two one-layer transformers\nTo compare features from different models, we need model-independent ways to represent a feature. \nOne natural approach is to think of a feature as a function assigning values to datapoints; two features would be similar in this sense if they take similar values over a diverse set of data. This general approach has been explored by a number of prior papers (e.g. ). In practice, this can be approximated by representing the feature as a vector, with indices corresponding to a fixed set of data points. We call the correlations between these vectors the activation similarity between features.\nA second natural approach is to think of a feature in terms of its downstream effects; two features would be similar in this sense if their activation changes their models' predictions in similar ways. In our one-layer model, a simple approximation to this is the logit weights. This approximation represents each feature as a vector with indices corresponding to vocabulary tokens. We call the correlations between these vectors the logit weight similarity between features.\nThese two notions of similarity correspond to the correlations of the points in the two scatter plots we used when [analyzing individual features](https://transformer-circuits.pub/2023/monosemantic-features/#feature-ananlysis) earlier. We've reproduced the plots for the Arabic feature below:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nFor each feature in run [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html), we find the closest feature by activation similarity in run [B/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html), which is a different dictionary learning run trained on different activations from a different transformer with different random seeds but otherwise identical hyperparameters. We find that many features are highly similar between models, with features in [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html) having a median activation correlation of 0.72 with the most similar feature from [B/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html). (We perform the same analysis finding the closest neurons between the transformers, and find significantly less similarity, with median activation correlation 0.46.) The features with low activation correlation between models may represent different \"feature splittings\" in the dictionaries learned or different \"true features\" learned by the base models.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nA natural next question is whether features that fire on the same tokens also have the same logit effects. That is, how well do activation similarity and logit weight similarity agree? \nSome gap between the two is visible for the Arabic feature above: the \"important tokens\" for the features' effects (the ones in Arabic script) are upweighted by features from both models, but there is a large cloud of tokens with smaller effects that appear to almost be isotropic noise, resulting in a logit weight correlation of just 0.23, significantly below the activation correlation of 0.91. \nIn the scatterplot below, we find that this kind of disagreement is widespread.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe most dramatic example of this disparity is for the features [A/1/3949](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3949) and [B/1/3321](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-3321), with an activation correlation of 0.98 but a negative logit weight correlation. These features fire on `pone` (and occasionally on `pgen` and `pcbi`) as abbreviations for the journal name PLOSOne in citations, like `@pone.0082392`, and predict the `.` that follows.The feature is bimodal, monosemantic in the larger of the modes, and fires on 0.02% of tokens. This means that at least 1 in 10,000 tokens in the Pile dataset are abbreviations for PLoS journals in citations! This is an example of how inspecting features can reveal properties of the dataset, in this case the strong bias of the Pile towards scientific content.\nZooming in on the logit weight scatterplot (inset in the figure above), we see that only the `.` token has high logit weight in both models, and that every other token is in the 'interference' portion of the logit weight distribution. Indeed, the model may simply not care about what the feature does to tokens which were already implausible because they are suppressed by the direct path, attention layer, or other features of the MLP. \nWe want to measure something more like \"the actual effect a feature has on token probabilities.\" One way to get at this would be to compute a vector of ablation effects for every feature on every data point; pairs of features whose ablations hurt the model's predictions on the same tokens must have been predicting the same thing. Unfortunately, this would be rather expensive computationally. Instead, we scale the activation vector of a feature by the logit weights of the tokens that empirically come next in the dataset to produce an attribution vector.Suppose feature f_i has logit weights v_{ik} for k \\in \\\\{1,\\ldots,n_{\\text{vocab}}\\\\}. At a given token t_j, we compute the activation of the feature f_i(t_j) and multiply it by the logit weight v_{it_{j+1}} of the token t_{j+1} that comes next to get an attribution score of f_i(t_j)v_{it_{j+1}}. The attribution vector is given by stacking the attribution scores for a random sampling of datapoints. This approximates the classic attribution method of multiplying the gradient by the activation, differing in that we ignore the denominators of the softmax and the layer norm. Correlations between those vectors provide an attribution similarity that combines both the activity of the feature with the effect it has on the loss. We find that the attribution similarity correlates quite highly with the activation similarity, meaning that features that were coactive between models were useful at predicting the same tokens.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nIn light of this, we feel that the activation correlation used throughout the paper is in fact a good proxy for both notions of universality in the context of our one-layer models.\n#### [Comparing features with the literature](https://transformer-circuits.pub/2023/monosemantic-features/#universality-literature)\nSo far, we've established that many of our features are universal in a limited sense. Features found in one of our transformers can also be found in an alternative version trained with a different random seed. But this second model has an identical architecture and was trained on identical data. This is the most minimal version of universality one could hope for. Despite this, we believe that many of the features we've found are universal in a deeper sense, because very similar features have been reported in the literature before.\nThe first comparison which struck us is that many features seem quite similar to neurons we previously found in one-layer SoLU models, which use an activation function designed to make neurons more monosemantic . In particular, we observed a base64 neuron, hexadecimal neuron, and all caps neuron in our SoLU investigations, and base64 ([A/0/45](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-45)), hexadecimal ([A/0/119](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-119)), and all caps ([A/0/317](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-317)) features here. Discussion of some of these neurons can be found in [Section 6.3.1](https://transformer-circuits.pub/2022/solu/index.html#section-6-3-1) of the SoLU paper.\nWe also find many features similar to Smith , who applies dictionary learning to the residual stream. In addition to us also observing preponderance of single token features they note (see [our interpretation](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs) of this phenomenon), we find a similar German detector (e.g. [A/0/493](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-493)) and similar title case detectors (e.g. [A/0/508](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-508)). Likewise, we find a number of features similar to Gurnee et al. , including a \"prime factors\" feature ([A/4/22414](https://transformer-circuits.pub/2023/monosemantic-features/vis/a4.html#feature-22414)) and a French feature ([A/0/14](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-14)).\nAt a more abstract level, many features we find seem similar to features reported in multimodal models by Goh et al. . For example, we find many similar features including an Australia feature ([A/3/16085](https://transformer-circuits.pub/2023/monosemantic-features/vis/a3.html#feature-16085)), Canada feature ([A/3/13683](https://transformer-circuits.pub/2023/monosemantic-features/vis/a3.html#feature-13683)), Africa feature ([A/3/14490](https://transformer-circuits.pub/2023/monosemantic-features/vis/a3.html#feature-14490)), and Israel-Palestine feature ([A/3/739](https://transformer-circuits.pub/2023/monosemantic-features/vis/a3.html#feature-739)) which predict locations in those regions when grammatically appropriate. This vaguely mirrors \"region neurons\" reported by Goh et al.'s paper. For other families of features, the parallels are less clear. For example, one of the most striking results of Goh et al. was person detector neurons (similar to famous results in neuroscience). We find some features that are person detectors in very narrow contexts, such as responding to a person’s name and predicting appropriate next words, or predicting their name (e.g. [A/1/3240](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3240) is somewhat similar to Goh et al.'s Trump neuron), but they seem quite narrow. We also don't find features that seem clearly analogous to Goh et al.'s emotion neurons.\n### [\"Finite State Automata\"](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-fsa)\nOne of the most striking phenomena we've observed in our study of the features in one-layer models is the existence of \"finite state automata\"-like assemblies of features. These assemblies aren't circuits in the conventional sense – they're formed by one feature increasing the probability of tokens, which in turn cause another feature to fire on the next step, and so on.The \"finite state automata\"-esque feature assemblies are also different from circuits in that the model didn't learn them to work together. Rather, in the course of learning to autoregressively model text, it learned features that interact via the token stream because of patterns in the real datasets. In contrast, a language model trained with reinforcement learning might have systems like this – circuits whose feature components interact via generated tokens – which co-evolved and adapted to work together during RL training.\nThe simplest example of this is features which excite themselves on the next token, forming a single node loop. For example, a base64 feature increases the probability of tokens like `Qg` and `zA` – plausible continuations which would continue to activate it.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nIt's worth noting that these examples are from [A/0](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html), a dictionary learning run which is not overcomplete (the dictionary dimensionality is 512, equal to the transformer MLP dimension). As we move to runs with larger numbers of features, the central feature will experience feature splitting, and become a more complex system.\nLet's now consider a two-node system for producing variables in \"all caps snake case\" (e.g. `ARRAY_MAX_VALUE`). One node ([A/0/207](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-207)) activates on the all caps text tokens, the other ([A/0/358](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-358)) on underscores:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThis type of two-node system is quite common for languages where Unicode characters are sometimes split into two tokens. (Again, with more feature splitting, these would expand into more complex systems.)\nFor example, Tamil Unicode characters ([block U+0B80–U+0BFF](https://www.compart.com/en/unicode/block/U+0B80)) are typically split into two tokens. For example, the character \"ண\" ([U+0BA3](https://www.compart.com/en/unicode/U+0BA3)) is tokenized as `\\xe0\\xae` followed by `\\xa3`. The first part (`\\xe0\\xae` or `\\xe0\\xaf`) roughly specifies the Unicode block, while the second component specifies the character within that block. Thus, it's natural for the model to alternate between two features, one for the Unicode prefix token, and one for the suffix token.\nA more complex example is Chinese. While many common Chinese characters get dedicated tokens, many others are split. This is further complicated by Chinese characters being spread over many Unicode blocks, and those blocks being large and cutting across many logical blocks specified in terms of bytes. To understand the state machine the model implements to handle this, the key observation is that complete characters are similar to the \"suffix\" part of a split character: both can be followed by either a new complete character, or a new prefix. Thus, we observe two features, one of which fires on either complete characters or the suffix (predicting either a new complete character, or a prefix), while the other only fires on the prefixes and predicts suffixes.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nLet's now consider a very simple four node system which models HTML. The \"main path\" through it is:\n  * [A/0/20](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-20) fires on open tags and predicts tag names\n  * [A/0/0](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-0) fires on tag names and predicts tag closes\n  * [A/0/30](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-30) fires on tag closes and predicts whitespace\n  * [A/0/494](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-494) fires on whitespace and predicts new tag opens.\n\n\nA prototypical sample this might generate is something like `<div>\\n\\t\\t<span>`.\nThe full system can be seen below:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nKeep in mind that we're focusing on the [A/0](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html) features where this is very simple – if we looked at [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html), we'd find something much more complex! One particularly striking shortcoming of the [A/0](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html) features is that they don't describe what happens when [A/0/0](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-0) emits a token like ` href`, which leads to a more complex state.\nIt's important to note that these features can be quite contextual. There are several features related to IRC transcripts which form a totally different finite state automata like system:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nA prototypical sample this might generate is something like `<nickonia_> lol ubuntu ;)`. Presumably the Pile dataset heavily represents IRC transcripts about linux.\nOne particularly interesting behavior is the apparent memorization of specific phrases. This can be observed only in runs with relatively large numbers of features (like [A/4](https://transformer-circuits.pub/2023/monosemantic-features/vis/a4.html)). In the following example, a sequence of features seem to functionally memorize the bolded part of the phrase `MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE`. This is a relatively standard legal language, and notably occurs in the file headers for popular open source software licenses, meaning the model likely saw it many times during training.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThis seems like an example of the mechanistic theory of memorization we described in Henighan et al.  – we observe features which appear to be relatively binary and respond to a very specific situation. This might also be seen as an instance of mechanistic anomaly detection : the model behaves differently in a specific, narrow case. It's somewhat surprising that something so narrow can be found in a model with only 512 neurons; from this perspective it's an interesting example of superposition's ability to embed many things in few neurons. On the other hand, because these mechanisms are buried deep in superposition, they are likely very noisy.\n  \n  \n  \n\n* * *\n  \n  \n\n## [Related Work](https://transformer-circuits.pub/2023/monosemantic-features/#related-work)\nSuperposition and attempts to resolve it have deep connections to many lines of research, including general investigations of interpretable features, linear probing, compressed sensing, dictionary learning and sparse coding, theories of neural coding, distributed representations, mathematical frames, vector symbolic architectures, and much more. Rather than attempt to do justice to all these connections here, we refer readers to the [related work section](https://transformer-circuits.pub/2022/toy_model/index.html#related) of Toy Models of Superposition where we discuss these topics in depth, and also to our essay [Distributed Representations: Composition & Superposition](https://transformer-circuits.pub/2023/superposition-composition/index.html) . Instead, we'll focus our discussion on work connected to our attempts to solve superposition, and also more recent advancements in our understanding of superposition.\n#### Superposition\nSince we published Toy Models of Superposition there has been significant further work attempting to better understand superposition. We briefly summarize below.\nIs superposition real? Gurnee et al. demonstrate some compelling examples of features which may be in superposition, using sparse linear probes. Separately, an exchange between Li et al. and Nanda et al. seems like an update on whether the general picture of features as directions is correct; Li et al. seemed to show that it wasn't, putting the hypothesis in jeopardy, which was then resolved by Nanda et al. \nWhen and why does superposition occur? Scherlis et al. provide a mathematical framework for thinking about monosemanticity vs polysemanticity. lsgos explored the effects of dropout on toy models of superposition.\nMemorization – In Henighan et al. , we studied the same toy model as in Toy Models of Superposition, but this time trained on many repetitions of finite-sized datasets. We found that small datasets are memorized in superposition, instead of generalizing features in the case of large datasets. Hobbhahn replicated some of these findings, and further showed extensions to other settings including bottlenecks between layers (analogous to the residual stream between MLP layers). Subsequently, in a monthly update we [examined](https://transformer-circuits.pub/2023/july-update/index.html#finite-data) the boundary between the memorization and generalization regimes and found a sharp phase transition, as well as dataset clustering in superposition on the memorization side of the boundary.\n#### Disentanglement and Architectural Approaches\nThere is also a rich and related literature on disentanglement, which seeks to find representations of data that separate out (disentangle) conceptually-distinct phenomena influencing the data. In contrast to superposition, this work typically seeks to find a number of factors of variation or features which are equal to the dimensionality of the space being represented, whereas superposition seeks to find more.\nThis is often approached as an architecture/training-time problem. For instance, Kim & Mnih proposed a method that pushes variational autoencoders to disentangle factors by encouraging independence across dimensions. Similarly, Chen et al. developed a Generative Adversarial Network approach that attempts to disentangle factors by maximizing the mutual information between a small subset of factors and the dataset. And Makhzani & Frey use a TopK activation to encourage sparsity and hence disentanglement. See Bengio et al. and Räuker et al. for a discussion of other such approaches.\nFramed this way, some architectural approaches to superposition may also be understood as attempts at disentanglement. For instance, in Elhage et al. , we proposed the SoLU activation function, which increases the number of interpretable neurons in transformers by encouraging features to align to the neuron basis. Unfortunately, it appears that training models with the SoLU activation function may make some neurons more interpretable at the cost of making others even less interpretable than before.\nSimilarly, Jermyn et al. studied MLP layers trained on a compressed sensing task and found multiple equal-loss minima, with some strongly polysemantic and others strongly monosemantic. This suggested that training interventions could steer models towards more monosemantic minima, though subsequent investigations on more realistic tasks suggested that the equal-loss property was specific to the chosen task.\nThese two examples of attempts to tackle superposition through architecture, and the challenges they encountered, highlight a key distinction between the problems of disentanglement and that of superposition: disentanglement fundamentally seeks to ensure that the dimensions in the model’s latent space are disentangled, whereas superposition hypothesizes that this disentanglement typically hurts performance (since success would require throwing away many features), and that models will typically respond to disentangling interventions by making some features more strongly entangled (as was found by both Mahinpei et al. and Elhage et al. 2022 in somewhat different contexts).\n#### Dictionary Learning and Features\nOur work builds on a longer tradition of using dictionary learning and sparse autoencoders to decompose neural network activations.\nEarly work in this space focused on word embeddings and other non-transformer neural networks. Faruqui et al. and [Arora ](https://arxiv.org/abs/1601.03764)[et al.](https://arxiv.org/abs/1601.03764) both found linear structure in word embeddings using sparse coding approaches. Subramanian et al. similarly found linear factors for word embeddings, in this case using a sparse autoencoder. Zhang et al. solved a similar problem using methods from dictionary learning while Panigrahi et al. approached this with Latent Dirichlet Allocation.\nMore recently, a number of works have applied dictionary learning methods to transformer models. Yun et al. applied dictionary learning to the residual stream of a 12-layer transformer to find an undercomplete basis of features.\nAt this point, our work in Toy Models advocated for dictionary learning as a potential approach to superposition. This motivated a parallel investigation by our colleagues Cunningham et al., published as a series of interim reports with very similar themes to this paper, culminating in a manuscript . We've been excited to see so many corroborating findings between our work.\nIn their interim reports, Sharkey et al. used sparse autoencoders to perform dictionary learning on a one-layer transformer, identifying a large (overcomplete) basis of features. (Sharkey et al. deserve credit for focusing on dictionary learning and especially the sparse autoencoder approach, while our investigation was only exploring it as one of several approaches in parallel.) This work was then partially replicated by Cunningham & Smith and Huben . Next, Smith used an autoencoder to find features in one MLP layer of a six-layer model. The resulting features appear interpretable, e.g. detecting ‘$’ in the context of LaTeX equations. In follow up work, Smith then extended this approach to the residual stream of the same model, identifying a number of interesting features (see [earlier discussion](https://transformer-circuits.pub/2023/monosemantic-features/#universality-literature)). Building on these results, Cunningham applied autointerpretability techniques from Bills et al. to features in the residual stream and an MLP layer of the same six-layer model, finding that the features discovered by the sparse autoencoder are substantially more interpretable than neurons.\n  \n  \n  \n\n* * *\n  \n  \n\n## [Discussion](https://transformer-circuits.pub/2023/monosemantic-features/#discussion)\n### [Theories of Superposition](https://transformer-circuits.pub/2023/monosemantic-features/#discussion-superposition)\nComing into this work, our understanding of superposition was mostly informed by Toy Models . This gave us a picture one might call the isotropic superposition model. Features are discrete, one-dimensional objects which repel from each other due to interference, creating a roughly evenly spaced organization of feature directions.\nThis work has persuaded us that our previous model was missing something crucial. At a minimum, features seem to clump together in higher density groups of related features. One explanation for this (considered briefly by Toy Models) is that the features may have correlated activations – firing together. Another – which we suspect to be more central – is that the features produce similar actions. The feature which fires on single digits in base64 predicts approximately the same set of tokens as the feature firing on other characters in base64, with the exception of other digits; these similar downstream effects manifest as geometrically close feature directions.\nMoreover, it isn't clear that features need to be one-dimensional objects (encoding only some intensity). In principle, it seems possible to have higher-dimensional \"feature manifolds\" (see earlier discussion [here](https://transformer-circuits.pub/2023/may-update/index.html#feature-manifolds)).\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThese hypotheses are not mutually exclusive. The convex hull of several correlated features might be understood as a feature manifold. On the other hand, some manifolds would not admit a unique description in terms of a finite number of one-dimensional features. (Perhaps this accounts for the continued [feature splitting](https://transformer-circuits.pub/2023/monosemantic-features/#phenomenology-feature-splitting) observed above.)\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nNevertheless, these experiments have left us more confident that some version of the superposition hypothesis (and the linear representation hypothesis) is true. The number of interpretable features found, the way activation level seems to correspond to \"intensity\" or \"confidence,\" the fact that logit weights mostly make sense, and the observation of \"interference weights\": all of these observations are what you would expect from superposition.\nFinally, we note that in some of these expanded theories of superposition, finding the \"correct number of features\" may not be well-posed. In others, there is a true number of features, but getting it exactly right is less essential because we \"fail gracefully\", observing the \"true features\" at resolutions of different granularity as we increase the number of learned features in the autoencoder.\n### [Are \"Token in Context\" Features Real?](https://transformer-circuits.pub/2023/monosemantic-features/#discussion-token-in-context)\nOne of the most common motifs we found were \"token-in-context\" features. They also represent many of the features that emerge via feature splitting with increasing dictionary size. Some of these are intuitive – borrowing an example from , it makes sense to represent \"die\" in German (where it's the definite article) as distinct from \"die\" in English (where it means \"death\" or \"dice\").\nBut why do we see hundreds of different features for \"the\" (such as \"the\" in Physics, as distinct from \"the\" in mathematics)? We also observe this for other common words (e.g. \"a\", \"of\"), and for punctuation like periods. These features are not what we expected to find when we set out to investigate one-layer models!\nTo make the question a bit more precise, it is helpful to borrow the language and examples of local vs compositional representations . Individually representing token-context pairs (such as \"the\" in Physics) is technically a [\"local code\"](https://transformer-circuits.pub/2023/superposition-composition/index.html#distributed-local). The more intuitive way to represent this would instead be a [\"compositional code\"](https://transformer-circuits.pub/2023/superposition-composition/index.html#distributed-compositional)[ – representing ](https://transformer-circuits.pub/2023/superposition-composition/index.html#distributed-compositional)\"the\" as an independent feature from Physics. So the thing we really want to ask is why we're observing a local code, and whether it's really what's going on. There are two hypotheses:\n  * The underlying transformer uses a compositional code, and a quirk of our dictionary learning scheme produces features using a local code.\n  * The underlying transformer is genuinely using a local code (at least in part), and dictionary learning is correctly representing this.\n\n\nIf the former holds, then better dictionary learning schemes may help uncover a more compositional set of features from the same transformer. Local codes are sparser than compositional codes, and our L1 penalty may be pushing the model too far towards sparsity.\nHowever, we believe the second hypothesis is likely to hold to some extent. Let's consider the example of \"the\" in Physics again, which predicts noun phrases in Physics: if the model represented \"the\" and Physics context independently, it would be forced to have logits be the sum of \"upweight tokens which come after the\" and \"upweight tokens which occur in Physics\". But the model might wish to have \"sharper\" predictions than this, which is only possible with a local code.\n### [Future Work](https://transformer-circuits.pub/2023/monosemantic-features/#discussion-future-work)\nScaling Sparse Autoencoders. Scaling the application of sparse autoencoders to frontier models strikes us as one of the most important questions going forward. We're quite hopeful that these or similar methods will work – Cunningham et al.'s work seems to suggest this approach can work on somewhat larger models, and we have preliminary results that point in the same direction. However, there are significant computational challenges to be overcome. Consider an autoencoder with a 100× expansion factor applied to the activations of a single MLP layer of width 10,000: it would have ~20 billion parameters. Additionally, many of these features are likely quite rare, potentially requiring the autoencoder to be trained on a substantial fraction of the large model's training corpus. So it seems plausible that training the autoencoder could become very expensive, potentially even more expensive than the original model. We remain optimistic, however, and there is a silver lining – it increasingly seems like a large chunk of the mechanistic interpretability agenda will now turn on succeeding at a difficult engineering and scaling problem, which frontier AI labs have significant expertise in.\nScaling Laws for Dictionary Learning. It's worth noting that there's enormous uncertainty about the dynamics of scaling dictionary learning and sparse autoencoders discussed above. As we make the subject model bigger, how does the ideal expansion factor change? (Does it stay constant?) How does the necessary amount of data change? The resolution of these questions will determine whether it's possible for this approach, if executed well, to scale up to frontier models. Ideally, we'd like to have scaling laws which could answer this.\nHow Can We Recognize Good Features? One of the greatest challenges of this work is that we're \"wandering in the dark\" to some extent. We don't have a great, systematic way to know if we're successfully extracting high quality features. Automated interpretability seems like a strong contender for solving this question. Alternatively, one might hope for some purely abstract definition (e.g. the [information-based metric](https://transformer-circuits.pub/2023/may-update/index.html#simple-factorization) proposal), but we have not yet seen compelling signs of life for this on real data. It would also be helpful to have metrics beyond MMCS, activation similarity, and attribution similarity for comparing sets of features for the purposes of assessing consistency and universality.\nScalability of Analysis. Suppose that sparse autoencoders fully solve superposition. Do we have a home run to fully mechanistically understanding models? It seems clear that there would be at least one other fundamental barrier: scaling analysis of models, so that we can turn microscopic insights into a more macroscopic understanding. Again, one approach here could be automated interpretability. But delegating the understanding of AI to AI may not be fully satisfying, for various reasons. It is possible that there may be other paths based on discovering larger scale structure (see discussion [here](https://transformer-circuits.pub/2023/interpretability-dreams/index.html#larger-scale)).\nAlgorithmic Improvements for Sparse Autoencoders. New algorithms refining the sparse autoencoder approach could be useful. One might explore the use of variational autoencoders (e.g., ), or sparsity promoting priors regularization techniques beyond a simple L1 penalty on activations (e.g., ), for example encouraging sparsity in the interactions between learned features in different layers. Earlier research has shown that noise injection can also increase neuron interpretability separately from an L1 penalty . \nAttentional Superposition? Many of the motivations for the presence of superposition in MLP layers apply to self-attention layers as well. It seems conceivable that similar methods may extract useful structure from attention layers, although a clear example has not yet been established (e.g., see our [May](https://transformer-circuits.pub/2023/may-update/index.html#attention-superposition) and [July](https://transformer-circuits.pub/2023/july-update/index.html#attn-skip-trigram) Updates). If this is true, addressing this may become a future bottleneck for the mechanistic interpretability agenda.\nTheory of Superposition and Features. Many fundamental questions remain for our understanding of superposition, even if the hypothesis is right in some very broad sense. For example, as discussed above, this work suggests extensions of the superposition hypothesis covering clusters of features with similar effects, or continuous families of features. We believe there is important work to be done in exploring the theory of superposition further, perhaps through the use of toy models.\n  \n  \n  \n\n* * *\n  \n  \n\n## [Comments & Replications](https://transformer-circuits.pub/2023/monosemantic-features/#comments)\nInspired by the original [Circuits Thread](https://distill.pub/2020/circuits/) and [Distill's Discussion Article experiment](https://distill.pub/2019/advex-bugs-discussion/), the authors invited several external researchers who we had previously discussed our preliminary results with to comment on this work. Their comments are included below.\n### [Replication & Tutorial](https://transformer-circuits.pub/2023/monosemantic-features/#comment-nanda)\nNeel Nanda is an external mechanistic interpretability researcher. This is a summary of [a blog post](https://www.alignmentforum.org/posts/fKuugaxt2XLTkASkk/open-source-replication-and-commentary-on-anthropic-s) replicating and extending this paper, with [an accompanying tutorial](https://colab.research.google.com/drive/1u8larhpxy8w4mMsJiSBddNOzFGj7_RTn?usp=sharing) to load some trained autoencoders and practice interpreting a feature. \nThe core results of this paper seem to replicate. I trained a sparse autoencoder on the MLP layer of an open source 1 layer GELU language model, and a significant fraction of the latent space features were interpretable.\nI've open sourced two trained autoencoders and [a tutorial for how to use them, and how to interpret a feature](https://colab.research.google.com/drive/1u8larhpxy8w4mMsJiSBddNOzFGj7_RTn?usp=sharing). I've also open sourced a (very!) rough [training codebase](https://github.com/neelnanda-io/1L-Sparse-Autoencoder), along with some [of the implementation details](https://www.alignmentforum.org/posts/fKuugaxt2XLTkASkk/open-source-replication-and-commentary-on-anthropic-s#Implementation_Details) that came up, and tips for training your own.\nI investigated [how sparse the decoder weights are in the neuron basis](https://www.alignmentforum.org/posts/fKuugaxt2XLTkASkk/open-source-replication-and-commentary-on-anthropic-s#Exploring_Neuron_Sparsity) and find that they’re highly distributed, with 4% well explained by a single neuron, 4% well explained by 2 to 10, and the remaining 92% dense. I find this pretty surprising! Despite this, kurtosis shows the neuron basis is still privileged.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nI also exhibit [some case studies](https://www.alignmentforum.org/posts/fKuugaxt2XLTkASkk/open-source-replication-and-commentary-on-anthropic-s#Case_Studies) of the features I found, like a title case feature, and an \"and I\" feature\nI didn’t find any dead features, but more than half of the features form [an ultra-low frequency cluster](https://www.alignmentforum.org/posts/fKuugaxt2XLTkASkk/open-source-replication-and-commentary-on-anthropic-s#Ultra_Low_Frequency_Features_Are_All_The_Same_Feature) (frequency less than 1e-4). Surprisingly, I find that this cluster is almost all the same feature (in terms of encoder weights, but not in terms of decoder weights). On one input 95% of these ultra rare features fired!\n  * The same direction forms across random seeds, suggesting it's a true thing about the model and not just an autoencoder artifact\n  * I failed to interpret what this shared direction was\n  * I tried to fix the problem by training an autoencoder to be orthogonal to this direction but it still forms an ultra-low frequency cluster (which all cluster in a new direction)\n\n\nOne question from this work is whether the encoder and decoder should be tied. I find that, empirically, [the decoder and encoder weights for each feature are moderately different](https://www.alignmentforum.org/posts/fKuugaxt2XLTkASkk/open-source-replication-and-commentary-on-anthropic-s#Misc_Questions), with median cosine similarity of only 0.5, which is empirical evidence they're doing different things and should not be tied. Conceptually, the encoder and decoder are doing different things: the encoder is detecting, finding the optimal direction to project onto to detect the feature, minimising interference with other similar features, while the decoder is trying to represent the feature, and tries to approximate the “true” feature direction regardless of any interference.\n### [Author Contributions Statement](https://transformer-circuits.pub/2023/monosemantic-features/#author-contributions)\n#### Infrastructure, Tooling, and Core Algorithmic Work\nGeneral Infrastructure – The basic framework for our dictionary learning work was built and maintained by Adly Templeton, Trenton Bricken, Tom Henighan, and Tristan Hume. Adly Templeton, in collaboration with Trenton Bricken and Tom Conerly, performed the engineering to scale up our experiments to large numbers of tokens. Robert Lasenby created tooling that allowed us to train extremely small language models. Yifan Wu imported the \"Pile\" dataset so we could train models on a standard external dataset. Tom Conerly broadly assisted with our infrastructure and maintaining high code quality. Adly Templeton implemented automatic plots comparing different experiments in a scan.\nSparse Autoencoders (Algorithms / ML) – Trenton initially implemented and advocated for sparse autoencoders as an approach to dictionary learning. Trenton Bricken and Adly Templeton then collaborated on the research needed to achieve our results, including scanning hyperparameters, iterating on algorithms, introducing the \"neuron resampling\" method, and characterizing the importance of dataset scale. (It's hard for the other authors to communicate just how much work Trenton Bricken and Adly Templeton put into iterating on algorithms and hyperparameters.) Josh Batson assisted by analyzing runs and designing metrics with Chris Olah, Adly Templeton and Trenton Bricken to measure success.\nAnalysis Infrastructure – The tooling to collect data for the visualizations was created by Trenton Bricken and Adly Templeton. Tom Conerly greatly accelerated this. Adly Templeton implemented the tooling to perform large scale feature ablation analysis, as well as creating the shuffled weights models as baselines.\nInterface – The interface for visualizing and exploring features was created by Brian Chen, with support from Shan Carter. It replaced a much earlier version by Trenton Bricken.\n#### Analysis\nFeature Deep Dives – The dataset examples and ablations came from Brian Chen's interface work and Adly Templeton's ablation infrastructure. Chris Olah developed the activation spectrum plots, broken down by proxy. Josh Batson developed the logit scatter plot visualization and the sensitivity analysis. Tom Henighan created the pinned sampling visualizations. Nick Turner, Josh Batson and Tom Henighan explored how best to analyze neuron alignment. Many people contributed to a push to systematically analyze many features in detail, including Tom Henighan, Josh Batson, Trenton Bricken, Adly Templeton, Nick Turner, Brian Chen, and Chris Olah.\nManual Analysis of Features – Nick Turner and Adam Jermyn created our rubric for evaluating if a feature interval is interpretable. Adam Jermyn manually scored a random set of features.\nAutomated Interpretability – Our automated interpretability pipeline was created by Trenton Bricken, Cem Anil, Carson Denison, and Amanda Askell. Trenton Bricken ran the experiments using it to evaluate our features. This built on earlier explorations of automated interpretability by Shauna Kravec, and was only possible due to infrastructure contributions by Tim Maxwell, Nicholas Schiefer, and Nicholas Joseph.\nFeature Splitting – Josh Batson, Adam Jermyn, and Chris Olah analyzed feature splitting. Shan Carter produced the UMAPs of features.\nUniversality – Josh Batson analyzed universality, with engineering support from Trenton Bricken and Tom Henighan. \n\"Finite State Automata\" – Chris Olah analyzed \"Finite State Automata\".\n#### Paper Production\nWriting – The manuscript was primarily drafted by Chris Olah, Josh Batson, and Adam Jermyn, with extensive feedback and editing from all other authors. The detailed appendix on dictionary learning was drafted by Adly Templeton and Trenton Bricken. Josh Batson and Chris Olah managed the revision process in response to internal and external feedback.\nIllustration – Diagrams were created by Chris Olah, Josh Batson, Adam Jermyn, and Shan Carter, based on data generated by themselves and others.\n#### Development of Intuition & Negative Results\nEarly Dictionary Learning Experiments – Many of the authors (including Trenton Bricken, Adly Templeton, Tristan Hume, Tom Henighan, Adam Jermyn, Josh Batson, and Chris Olah) did experiments applying both more traditional dictionary learning methods and sparse autoencoders to a variety of toy problems and small MNIST models. Much of this work focused on finding metrics which we hoped would tell us whether dictionary learning had succeeded in a simple context. These experiments, along with theoretical work, built valuable intuition and helped us discover several algorithmic improvements, and clarified challenges with traditional dictionary learning.\nSparse Architectures – Brian Chen ran the experiments and generated the counter-examples which persuaded us that training models for sparse activations was less promising. Crucially, he produced models where neurons typically activated in isolation, and showed they were still polysemantic. He then produced the counter-example described in the text. Robert Lasenby implemented infrastructure that made it easier to experiment with sparse activation architectures.\n#### Other\nSupport – Alex Tamkin, Karina Nguyen, Brayden McLean, and Josiah E Burke made a variety of contributions through infrastructure, operational support, labeling features, and commenting on the draft.\nLeadership – Tom Henighan led the dictionary learning project. Tristan Hume led an early version of the project, before handing it over to Tom Henighan. Shan Carter managed the overall team. Chris Olah provided general research guidance.\n### [Acknowledgments](https://transformer-circuits.pub/2023/monosemantic-features/#acknowledgments)\nWe are deeply grateful to Martin Wattenberg, Neel Nanda, Hoagy Cunningham, David Lindner, Logan Smith, Steven Bills, William Saunders, Jonathan Marcus, Daniel Mossing, Nick Cammarata, Robert Huben, Aidan Ewart, Nicholas Sofroniew, Anna Golubeva, Bruno Olshausen for their detailed comments and feedback, which greatly improved our work.\nWe are also grateful to our colleagues at Anthropic who generously provided feedback, and many of whom also helped us explore the features discovered by dictionary learning as part of a \"Feature Party\". We particularly thank Oliver Rausch, Pujaa Rajan, Anna Chen, Alexander Silverstein, Marat Freytsis, Maryam Mortazavi, Mike Lambert, Justin Spahr-Summers, Emmanuel Ameisen, Andre Callahan, Shannon Yang, Zachary Witten, Zac Hatfield-Dodds, Karina Nguyen, Avital Balwit, Amanda Askell, Brayden McLean, and Nicholas Schiefer.\nOur work is only possible because of the extensive support of all our colleagues at Anthropic, from infrastructure, to engineering, to operations and more. It's impossible for us to list all the people whose work indirectly supported this paper, because there are so many, but we're deeply grateful for their support.\n### [Citation Information](https://transformer-circuits.pub/2023/monosemantic-features/#citation)\nPlease cite as:\n\n```\nBricken, et al., \"Towards Monosemanticity: Decomposing Language Models With Dictionary Learning\", Transformer Circuits Thread, 2023.\n```\n\nBibTeX Citation:\n\n```\n    @article{bricken2023monosemanticity,\n       title={Towards Monosemanticity: Decomposing Language Models With Dictionary Learning},\n       author={Bricken, Trenton and Templeton, Adly and Batson, Joshua and Chen, Brian and Jermyn, Adam and Conerly, Tom and Turner, Nick and Anil, Cem and Denison, Carson and Askell, Amanda and Lasenby, Robert and Wu, Yifan and Kravec, Shauna and Schiefer, Nicholas and Maxwell, Tim and Joseph, Nicholas and Hatfield-Dodds, Zac and Tamkin, Alex and Nguyen, Karina and McLean, Brayden and Burke, Josiah E and Hume, Tristan and Carter, Shan and Henighan, Tom and Olah, Christopher},\n       year={2023},\n       journal={Transformer Circuits Thread},\n       note={https://transformer-circuits.pub/2023/monosemantic-features/index.html}\n    }\n\n```\n\n### [Feature Proxies](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-proxies)\nIn order to analyze features, we construct \"proxies\" for each feature which can be easily computed. Our proxies are the log-likelihood ratio of a string under the feature hypothesis and under the full empirical distribution. For example, \\log(P(s|\\text{base64}) / P(s)) or \\log(P(s|\\text{Arabic Script}) / P(s)). We use log-likelihood proxies on the intuition that features, since they linearly interact with the logits, will be incentivized to track log-likelihoods.\nSince the activation of a feature at a token may be in part due to the preceding tokens (for example the Arabic feature could fire more strongly towards the end of a long Arabic string), we maximize the log-likelihood ratio over different prefixes leading up to and including the final token.\nBut how can we estimate the different terms? Let's consider each part separately.\n#### Estimating P(s)\nTo compute P(s), we use the unigram distribution over tokens and compute P(s) = \\prod_{t\\in s} p(t).\n#### Estimating P(s|\\text{base64}) (and other characters features)\nFor base64, we assume that we're modeling a random base64 string where the probability of a character c is P(c) = 1/64 if it's in a base64 character, and 10^{-10} otherwise. We then compute P(s) = \\prod_{c\\in s} P(c).\nWe model DNA as a random string of `[ATCG]`, each with probability 1/4.\n(Although we don't use them in this draft, we also found this approach helpful for hexadecimal, binary, and similar features.)\n#### Estimating P(s|\\text{Arabic}) (and other scripts features)\nWe also construct proxies for languages written in scripts that use distinctive unicode blocks. We'll use Arabic as an example of this.\nFor P(s|\\text{Arabic}), we use Bayes rule (i.e. P(s|\\text{Arabic}) = P(\\text{Arabic}|s) \\cdot P(s) / P(\\text{Arabic})). For P(s), we use the method discussed above. For the other terms, we exploit the fact that it is easy to detect whether a character is in a Unicode block associated with Arabic script (including e.g. [U+0600–U+06FF](https://www.compart.com/en/unicode/block/U+0600), [U+0750–U+077F](https://www.compart.com/en/unicode/block/U+0750)). Note this will count common characters shared between languages, like punctuation, as non-Arabic characters. If a string s consists entirely of Arabic characters, P(Arabic|s) is 1; if it contains non-Arabic characters, P(\\text{Arabic}|s) is assigned a tiny probability of 1e−10. (Keep in mind that we will be maximizing over multiple prefix strings.) If s consists of a single token which isn't a complete unicode character, we sample random occurrences of it and observe the fraction of the time it is used in a given script. We also estimate P(\\text{Arabic}) by applying this heuristic to a random sample of the dataset.\n### [Feature Ablations](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-ablations)\nWe perform feature ablations by running the model on an entire context up through the MLP layer, running the autoencoder to compute feature activations, subtracting the feature direction times its activation from the MLP activation on each token in the context (replacing \\mathbf{x}^j with \\mathbf{x}^j - f_i(\\mathbf{x}^j)\\mathbf{d}_j) and then completing the forward pass. We record the resulting change in the predicted log-likelihood of each token in the context in the color of an underline of that token. Thus if a feature were active on token [B] in the sequence [A][B][C], and ablating that feature reduced the odds placed on the prediction of C, then there would be an orange background on [B] (the activation) and a blue underline on [C] (the ablation effect), indicating that ablating that feature increased the model’s loss on the prediction of [C] and hence that feature is responsible for improving the model’s ability to predict [C] in that context.\n### [Feature Interpretability Rubric](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-rubric)\nOur rubric for scoring how interpretable features are has the following instructions:\n  1. Form an interpretation of the feature based on the positive logits and feature activations across all intervals.\n  2. On a scale of 0–3, rate your confidence in this interpretation.\n  3. On a scale of 0–5, rate how consistent the high-activation (bolded) tokens are with your interpretation from (1).\n  4. On a scale of 0–3, rate how consistent the positive logit effects are with your interpretation from (1).\n  5. If some of the positive logit effects were inconsistent with your interpretation, was there a separation in effect size between the consistent and inconsistent ones? If so, score as 1, otherwise or if not applicable, score as 0.\n  6. On a scale of 0–3, how specific is your interpretation of this feature?\n\n\nThe total score for a feature’s interpretability is the sum of ratings across these steps. Note that the maximum score is 14, because a perfect score on item #4 implies that item #5 is not applicable.\n### [Feature Density Histograms](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-density)\nWhile iterating on different techniques, one important proxy for autoencoder performance is feature density. Each feature in our autoencoder only activates on a very small percentage of the total tokens in the training set. We define the feature density of each feature as the fraction of tokens on which the feature has a nonzero value.\nWe hypothesize that language models contain a large number of features across a distribution of feature densities, and that lower-density features are harder for our autoencoder to discover because they appear less often in the training dataset. Using large training datasets was an attempt to recover such low-density features.\nAs one way to measure the performance of an autoencoder, we take all the feature densities and plot a histogram of their distribution on a log scale. Roughly, we look for two histogram metrics: the number of features we’ve recovered, and the minimum feature density among features that we’ve recovered. Anecdotally, we find that these metrics are a decent proxy for subjective autoencoder performance and useful for quick iteration cycles.\nIdeally, the autoencoder neurons would be either meaningful features or completely dead. Indeed, training does completely leave some autoencoder neurons unused. Many neurons, however, seem to fall into a third category, which we tentatively name the ultralow density cluster.\nMany of the feature density histograms are bimodal. For example, consider the feature density histogram for A/16:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThis histogram has two main modes: a left mode around 10^{-7} (corresponding to the ultralow density cluster) and a right mode around 10^{-5}.The right mode in this example may actually be split into two modes, one around 10^{-5} and another around 10^{-3}. We are still investigating the nature of this split. We call the right mode the “high density cluster”, although the absolute magnitude of the density is very low. Anecdotally, almost all of the features in the high density cluster are interpretable, but almost none of the features in the ultralow density cluster are. As an aside, we suspect that this bimodality in feature density might partially explain the bimodality in cross-dictionary MCS observed by Huben . \nIn many preliminary small-scale experiments, the two clusters are often perfectly separable. However, for runs with a large number of autoencoder neurons and a large L1 coefficient these two clusters often overlap.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nEven in these cases, the two clusters can be separated by combining other statistics. For each feature, in addition to density, we plot the bias and the dot product of the vectors in the decoder and the encoder corresponding to each feature. Along these three dimensions, the high density cluster is clearly separable.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nThe ultralow density cluster appears to be an artifact of the autoencoder training process and not a real property of the underlying transformer. We are continuing to investigate the source of this phenomenon and other ways to either mitigate these features or to robustly automatically detect and filter them.\nBelow, we present feature density histograms for all Seed A runs, grouped by L1 coefficient. As we increase the number of autoencoder neurons (or “N learned sparse”), we continue to discover not only more features but also rarer features, suggesting that we have not exhausted the features in our one-layer transformer.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n### [Transformer Training and Preprocessing](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-transformer)\nThe one-layer transformers we study are trained on the Pile . We chose to train these models on the Pile over Anthropic's internal dataset in order to make our experiments more reproducible. While we think many features are universal, others are very likely idiosyncratic to the dataset.\nWe train the transformers on 100 billion tokens using the Adam optimizer . We hypothesize that a very high number of training tokens may allow our model to learn cleaner representations in superposition. These transformers have a residual stream dimension of 128, and an inner MLP dimension of 512. The MLP activation is a ReLU.\n### [Advice for Training Sparse Autoencoders: Autoencoder Dataset and Basic Training](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-dataset)\nTo create the dataset for autoencoder training, we evaluate the transformers on 40 million contexts from the Pile and collect the MLP activation vectors after the ReLU for each token within each context. We then sample activation vectors from 250 tokens in each context and shuffle these together so that samples within a batch come from diverse contexts.\nFor training, we sample MLP activation vectors without replacement using a batch size of 8192. We’ve found that sampling without replacement (i.e., not repeating data) is important for optimal results. We perform 1 million update steps, using just over 8 billion activation vectors out of the total dataset size of 10 billion. We use an Adam optimizer to minimize the sum of mean squared error loss and an L1 regularization penalty on the hidden layer activations of the autoencoder.\n### [Advice for Training Sparse Autoencoders: Autoencoder Architecture](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder)\nOur dictionary learning model is a one hidden layer MLP. It is trained as an autoencoder, using the input weights as an encoder and output weights as the decoder. The hidden layer is much wider than the inputs and applies a ReLU non-linearity. We use the default Pytorch Kaiming Uniform initialization . \nFormally, let n be the input and output dimension and m be the autoencoder hidden layer dimension. Given encoder weights W_e \\in \\mathbb{R}^{m \\times n}, decoder weights W_d \\in \\mathbb{R}^{n \\times m} with columns of unit norm, and biases \\mathbf{b}_e \\in \\mathbb{R}^{m}, \\mathbf{b}_d \\in \\mathbb{R}^{n}, the operations and loss function over a dataset X are: \n\\begin{aligned} \\bar{\\mathbf{x}} &= \\mathbf{x} - \\mathbf{b}_{d} \\\\\\ \\mathbf{f} &= \\text{ReLU}( W_e \\bar{\\mathbf{x}}+\\mathbf{b}_e ) \\\\\\ \\hat{\\mathbf{x}} &= W_d \\mathbf{f}+\\mathbf{b}_d \\\\\\ \\mathcal{L} &= \\frac{1}{|X|} \\sum_{\\mathbf{x}\\in X} ||\\mathbf{x}-\\hat{\\mathbf{x}}||_2^2 + \\lambda||\\mathbf{f}||_1 \\end{aligned}\nNote that our hidden layer is overcomplete m\\geq n, and the MSE is a mean over each vector element while the L1 penalty is a sum (with \\lambda being the L1 coefficient). The hidden layer activations \\mathbf{f} are our learned features. As shown, we subtract the decoder bias from the inputs, and call this a pre-encoder bias. Recall that the decoder is also referred to as the “dictionary” while the encoder can be thought of as a linear, amortized approximation to more powerful sparse coding algorithms.\nOur modifications to a standard autoencoder are backed by theory and empirical ablations. However, because these modifications appeared at different periods during our research and can interact with each other, we do not provide rigorous ablation experiments and instead provide intuition and anecdotal evidence for why each is justified. There are likely many important modifications that could further improve performance.\nHere is a summary of the modifications sorted both between and within sections in rough order of importance \n  * Pre encoder bias – this boosted performance in toy models.\n  * Decoder weights not tied – the learned encoder weights are often far from the corresponding decoder weights and this is important for increasing model capacity.\n  * Neuron resampling – helps finding more features and achieve a lower total loss.\n  * Many training steps – beyond when the training loss looks to have plateaued.\n  * Learning rate sweep – lower learning rate results in more real features. \n  * Interaction Between Adam and Decoder Normalization – a principled way to account for how the dictionary vector L2 normalization interacts with Adam, resulting in lower total loss.\n\n\n#### [Pre-Encoder Bias](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-bias)\nTesting the autoencoder on toy problems with a bias, we found that it would fail to produce the “bounce plots” we [described previously](https://transformer-circuits.pub/2023/may-update/index.html#simple-factorization) unless we added a learned bias term to both the decoder output and the encoder input. We constrain the pre-encoder bias to equal the negative of the post-decoder bias and initialize it to the geometric median of the dataset. These investigations were motivated by the observations by Hobbhahn .\n#### [Decoder Weights Not Tied](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-untied)\nIt is common to tie the encoder and decoder weights of single hidden layer autoencoders . However, we find that in our trained models the learned encoder weights are not the transpose of the decoder weights and are cleverly offset to increase representational capacity. Specifically, we find that similar features which have closely related dictionary vectors have encoder weights that are offset so that they prevent crosstalk between the noisy feature inputs and confusion between the distinct features.\nFor example, in [\"Bug\" 2: Multiple Features for a Single Context](https://transformer-circuits.pub/2023/monosemantic-features/#features-seem-like-bugs-2) we described 3 features handling different types of base64 strings. These all have similar dictionary vectors, but it turns out that the encoder weight vectors, while still similar, are more tilted away to help distinguish them.\nOne way of viewing this is that our autoencoder is learning an amortized, linear approximation to a multi-step, non linear sparse coding algorithm. Allowing the weights of the encoder to be independent from the decoder enables more representational capacity as we empirically observe. \n#### [Neuron Resampling](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-resampling)\nOver the course of training, a subset of autoencoder neurons will have zero activity across a large number of datapoints. We find that “resampling” these dead neurons during training improves the number of likely-interpretable features (i.e., those in the high density cluster, see [Feature Density Histograms](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-density)) and reduces total loss. This resampling may be compatible with the Lottery Ticket Hypothesis and increase the number of chances the network has to find promising feature directions. \nAn interesting nuance around dead neurons involves the ultralow density cluster. We find that if we increase the number of training steps then networks will kill off more of these ultralow density neurons. This reinforces the use of the high density cluster as a useful metric because there can exist neurons that are de facto dead but will not appear to be when looking at the number of dead neurons alone.\nThe number of dead neurons that appear over training appears to depend upon a number of factors including but not limited to: learning rate (too high); batch size (too low); dataset redundancy (too many tokens per context or repeated epochs over the same dataset); number of training steps (too many); optimizer used. Our results here agree with those of Bricken et al. .\nBetter resampling strategies is an active research area of ours. The approach used in this work is as follows: \n  1. At training steps 25,000, 50,000, 75,000 and 100,000, identify which neurons have not fired in any of the previous 12,500 training steps. \n  2. Compute the loss for the current model on a random subset of 819,200 inputs.\n  3. Assign each input vector a probability of being picked that is proportional to the square of the autoencoder’s loss on that input.\n  4. For each dead neuron sample an input according to these probabilities. Renormalize the input vector to have unit L2 norm and set this to be the dictionary vector for the dead autoencoder neuron. \n  5. For the corresponding encoder vector, renormalize the input vector to equal the average norm of the encoder weights for alive neurons × 0.2. Set the corresponding encoder bias element to zero.\n  6. Reset the Adam optimizer parameters for every modified weight and bias term. \n\n\nThis resampling procedure is designed to seed new features to fit inputs where the current autoencoder performs worst. Resetting the encoder norm and bias are crucial to ensuring this resampled neuron will only fire weakly for inputs similar to the one used for its reinitialization. We do this to minimize interference with the rest of the network. \nThis resampling approach outperforms baselines including no resampling and reinitializing the relevant encoder and decoder weights using a default Kaiming Uniform initialization. However, there are likely better resampling methods to be developed – this approach still causes sudden loss spikes, and resampling too frequently causes training to diverge.\n#### [Learning Rate Sweep](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-learning-rate-sweep)\nWe performed several training runs using multiple learning rates while also varying the number of training steps and found that lower learning rates, when given sufficient training steps, result in lower total loss and more “real” features discovered (as indicated by the [Feature Density Histograms](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-density)). We also tried annealing our learning rate over the course of training but found that this did not further increase performance.\n#### [Interaction Between Adam and Decoder Normalization](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-autoencoder-optimization)\nRecall that we constrain our dictionary vectors to have unit norm. Our first naive implementation simply reset all vectors to unit norm after each gradient step. This means any gradient updates modifying the length of our vector are removed, creating a discrepancy between the gradient used by the Adam optimizer and the true gradient. We find that instead removing any gradient information parallel to our dictionary vectors before applying the gradient step results in a small but real reduction in total loss.\n### [Advice for Training Sparse Autoencoders: Hyperparameter Selection](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-hyperparameters)\nThe ultimate goal of dictionary learning is to produce features that are interpretable and also accurately represent the underlying model. Quantitatively measuring interpretability is currently difficult and slow, so we frequently look at proxy metrics when testing changes or optimizing hyperparameters. Our methods for optimizing hyperparameters are far from perfect, but we want to share insights that may be useful for other researchers doing similar work. \nWe most frequently look at the following proxy metrics: \n  * Training loss: This is the simplest metric, and useful for testing changes such as optimizers or learning rates. Unfortunately, it isn’t meaningful to compare training loss across hyperparameters that change the loss function, such as L1 coefficients.\n  * [Feature Density Histograms](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-feature-density): Specific metrics from these histograms include:\n\n\n  * The number of alive features outside of the ultralow density cluster\n  * The minimum feature density at which we see a significant number of non-ultralow-density-cluster features.\n  * The number of features with density above 1%. A significant number of features above this level seems to correspond to an L1 coefficient that is too low.\n\n\n  * \\mathbf{L^0} norm: The average number of nonzero entries in a sparse representation. We don’t yet know of a principled way to determine a target level of sparsity, but we generally target a L^0 norm that is less than 10 or 20. We especially distrust solutions where the L^0 norm is a significant fraction of the transformer’s activation dimensionality.\n  * Reconstructed Transformer NLL: We would like the features we discover to explain almost all of the behavior of the underlying transformer. One way to measure this is to take a transformer, run the MLP activations through our autoencoder, replace the MLP activations with the autoencoder predictions, measure the loss on the training dataset, and calculate the difference in loss. \n\n\nWe often normalize this by dividing by the difference in loss between the baseline transformer’s performance and its performance after ablating the MLP layer. This gives us a fraction of the MLP’s loss contribution that is explained by our transformer. However, the performance with an ablated MLP may be an especially bad baseline, so this percentage is considered an overestimate. Ad hoc experiments (data not shown) show that similarly high percentages can be reached by training a transformer from scratch with a small MLP.\nThe L^0 norm and Reconstructed Transformer NLL metrics, taken together, display a human-interpretable tradeoff between sparsity and explained behavior:\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nEach line on this plot shows data for a single L1 coefficient and varying autoencoder sizes. The lowest L1 coefficient is excluded to create a reasonable scale.\n### [Activation Visualization](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-visualization)\nVisualizations were produced using a 100 million token subset of our training dataset where only 10 tokens are taken from every context. We display a window of four tokens on each side of the selected token for visualization.   \n  \n\n### [Feature UMAPs](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-umap)\nWe construct two UMAP embeddings of learned features from our transformer. We represent each feature as a vector of dimension 512 (the number of neurons), its column in the decoder matrix W_d. The first UMAP embeds the features from A/0 and A/1 jointly into 2 dimensions, and uses hyperparameters `n_neighbors=15`, `metric=\"cosine\"`, `min_dist=0.05`. The second UMAP uses the same hyperparameters, but with `min_dist=0.01`, and also includes the features from A/2. We additionally organize the points in the first UMAP into clusters by applying HDBSCAN with `min_cluster_size=3` to a 10-dimensional UMAP embedding of the same features, fit with hyperparameters `n_neighbors=15`, `metric=\"cosine\"`, `min_dist=0.1`. The list of features is sorted by a one-dimensional UMAP fit to the unclustered features and the cluster mediods. \n### [Automated Interpretability](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-automated)\nHere we provide more details on the implementation of our automated interpretability before explaining importance scoring and providing additional results. \n#### [Experimental Setup](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-automated-setup)\nIn order to get feature explanations Claude 2 is provided with a total of 49 examples: ten examples from the top activations interval; two from the other 12 intervals; five completely random examples; and ten examples where the top activating tokens appear in different contexts.We are explicit in keeping all of these examples separate from the ones used when we test on activation predictions.For example, in the main text we discuss a feature that fires for token \"the\" in the context of machine learning, the ten examples here would contain the word \"the\" but across different contexts. We also provide the top and bottom logits and labels for what interval each example came from. In addition, we give eight human demonstrations on held out features that include chain of thought reasoning. Following Bills et al. all of our activations are quantized to be between 0–9 inclusive. Finally, we ask the model to be succinct in its answer and not provide specific examples of tokens it activates for. \nUsing the explanation generated, in a new interaction Claude is asked to predict activations for sixty examples: six from the top activations; two from the other 12 intervals; ten completely random; and twenty top activating tokens out of context. The same eight held out demonstration features show how each token should elicit a predicted activation. All of the examples are shuffled, the logits and interval each came from is removed, and blanks are left where the model should fill in its activation predictions. For the sake of computational efficiency, Claude scores all sixty examples in a single shot, repeating each token followed by its predicted activation. In an ideal setting, each example would be given independently as its own prompt.\nNote that there are instances where a feature has less than 49 explanation examples and 60 prediction examples because we remove any repeated examples that appear in the feature visualization intervals. However, this occurs for a negligible number of features and we drop any of those with fewer than 20 examples in total. \nFor the sake of comparison to Bills et al. the figure below includes Pearson correlation in addition to Spearman that is otherwise used throughout the text. The distributions and overall conclusions are almost identical to Spearman. \n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nWe also systematically catalog the feature output weight logits with Claude. We use the same explanation Claude generated to predict activations and provide 100 logits, 50 taken randomly and 50 which are the 10th–60th largest positive logits (the top 10 were used to generate the original explanations and thus held out). Claude was asked to output a binary label for if this logit would plausibly come after the feature fired, given the explanation. Note that the chance performance is 50%. The A/1 features are quite a bit better than the neurons at a median of 0.74 versus 0.53 for the neurons and 0.5 for the randomized transformer. \n#### [Randomized Transformer](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-automated-randomized)\nAs a further ablation, we compare our human and automated interpretability approaches to the randomized transformer. It is unsurprising that the logit weight predictions are centered around random chance for the randomized transformer by the very nature of its random weights. Manual interpretability also favors the A/1 features. However, for automated interpretability the randomized features have a higher median score.\nIt turns out that the randomized transformer features are distinctly bimodal, consisting of single token and polysemantic clusters which are separated out in the inset at the bottom of the following figure. \n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nFeatures are labeled as being \"single token\" if for all 20 examples in the top activation interval, the token with the largest activation is the same. The single token features of the randomized transformer are not only more numerous but also more truly single token, for example, firing for the same token even in the weakest activation intervals. This makes them more trivial to score as evidenced by the high scoring yellow peak (bottom row left side of the figure). \nMeanwhile, for the not \"single token\" cluster, the A/1 features score higher while the randomized features in fact have a similar distribution to that of the polysemantic neurons. \nNote that this bimodal clustering of the randomized transformer features is also supported by the manual human analysis where there is both a cluster at 0, the same score given to the polysemantic neurons in the main text, and a non zero cluster. The non-zero cluster scores worse than the A/1 Features here (unlike for the automated approach) because part of the rubric accounts for the logit interpretability, which scores zero in the shuffled case. \nThe fact that the randomized transformer learns more highly single token features fits with the hypothesis that dictionary learning will learn the features of the dataset rather than the representations of the transformer.\n#### [Importance Scoring](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-automated-importance-scoring)\nImportance scoring is the principled way to correct for the fact that we did not randomly sample all of the examples our model is asked to score. Formally, we assign to all tokens from random examples a weight of 1 and examples from each feature interval a weight of `feature_density*interval_probability`. `feature_density` is the fraction of times this feature fires over a large portion of the dataset and `interval_probability` converts the distribution of non-zero activations into a categorical distribution corresponding to each interval and uses the corresponding probability. We then use these weights in our Spearman correlation as before.The top activating tokens out of context examples inherit the weighting of the top activating examples quantile.\nThe issue with importance scoring is that because our features have densities typically in the range of 1e−3 to 1e−6 (see next Appendix section), almost all of the weight is assigned to the random examples for which features almost never fire. As a result, it is possible for features to score well by almost always predicting zero and in turn favors explanations that are overly restrictive that say what the feature does not fire for, instead of what it does fire for. Conversely, explanations are highly penalized for even a small false positive rate, since the true rate of occurrences is so low.\nThe figure below shows importance scores for features and neurons. Note that many of the features now have scores very close to zero while the neurons are relatively less affected. We hypothesize this is because the neurons, in being more polysemantic, have more non-zero activations that test their explanations on both things it does and doesn't fire for. Empirically, taking an average over features, 88% of their true activations from the random examples are zero versus 54% for the neurons (the medians are ~91% and 58%, respectively). \n![](https://transformer-circuits.pub/2023/monosemantic-features/)\n#### [Explanation Lengths Caveat](https://transformer-circuits.pub/2023/monosemantic-features/#appendix-automated-explanation-length)\nClaude can at times fail to wrap its explanation in the `<answer></answer>` explanation tags resulting in the full chain of thought being used instead. This can be much longer and leak more specific words that the feature fires for. \nHere we plot the number of characters in each explanation where the second mode to the right highlights these instances. However, not only does this happen infrequently, it also occurs across the features, neurons, and randomized transformer results. In fact, it happens the most often for neurons, occurring 19% of the time versus 15% for the randomized transformer and 13% for the features. Most importantly, it does not result in significantly different scores when we split out the performance of those with long explanations from short ones.\n![](https://transformer-circuits.pub/2023/monosemantic-features/)\nTo conclude, while automated interpretability is a difficult task for models that we have only begun working on, it has already been very useful for quickly understanding dictionary learning features in a scalable fashion. We encourage readers to explore the visualizations with automated interpretability scores: [A/1](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html), [A/0](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html), [randomized model features](https://transformer-circuits.pub/2023/monosemantic-features/vis/random1.html), [neurons](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html).\n### Footnotes\n  1. For more discussion of this point, see [Distributed Representations: Composition and Superposition](https://transformer-circuits.pub/2023/superposition-composition/index.html).[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-1)\n  2. We'd particularly highlight an exciting exchange between Li et al. and Nanda et al. : Li et al. found an apparent counterexample where features were not represented as directions, which was then resolved by Nanda finding an alternative interpretation in which features were directions.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-2)\n  3. While not the focus of this paper, we could also imagine decomposing other portions of the model’s latent state or activations in this way. For instance, we could apply this decomposition to the residual stream (see e.g. ), or to the keys and queries of attention heads, or to their outputs.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-3)\n  4. This property is closely related to the desiderata of Causality, Generality, and Purity discussed in Cammarata et al. , and those provide an example of how we might make this property concrete in a specific instance.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-4)\n  5. This is similar in spirit to the evidence provided by influence functions.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-5)\n  6. For the model discussed in this paper, we trained the autoencoder on 8 billion datapoints.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-6)\n  7. We tie the biases applied in the input and output, so the result is equivalent to subtracting a fixed bias from all activations and then using an autoencoder whose only bias is before the encoder activation.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-7)\n  8. Note that using an MSE loss avoids the challenges with polysemanticity that we discussed above in [Why Not Architectural Approaches?](https://transformer-circuits.pub/2023/monosemantic-features/#why-not-architectures)[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-8)\n  9. Note that this linear structure makes it even more likely that features should be linear. On the one hand, this means that the linear representation hypothesis is more likely to hold for this model. On the other hand, it potentially means that our results are less likely to generalize to multilayer models. Fortunately, others have studied multilayer transformers with sparse autoencoders and found interpretable linear features, which gives us more confidence that what we see in the one-layer model indeed generalizes .[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-9)\n  10. For example, [one neuron](https://transformer-circuits.pub/2023/monosemantic-features/vis/a-neurons.html?ordering=index#feature-83) in our transformer model responds to a mix of academic citations, English dialogue, HTTP requests, and Korean text. In vision models, there is a classic example of a neuron which responds to cat faces and fronts of cars .[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-10)\n  11. This kind of split tokenization is common for Unicode characters outside the Latin script and the most common characters from other Unicode blocks.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-11)\n  12. For example, in the figure above, there is a newline character ⏎ that our feature fires on but our proxy assigns a low score to because it is outside the Unicode block.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-12)\n  13. Why do we believe large activations have larger effects? In the case of a one-layer transformer like the one we consider in this paper, we can make a strong case for this: features have a linear effect on the logits (modulo rescaling by layer norm), and so a larger activation of the feature has a larger effect on the logits. In the case of larger models, this follows from a lot of conjectures and heuristic arguments (e.g. the abundance of linear pathways in the model and the idea of linear features at each layer), and must be true for sufficiently small activations by continuity, but doesn't have a watertight argument.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-13)\n  14. (also called “logit attribution”, see similar work e.g. )[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-14)\n  15. We exclude weights corresponding to extremely rare or never used vocabulary elements. These are perhaps similar to the \"anomalous tokens\" (e.g., \"SolidGoldMagikarp\") of Rumbelow & Watkins .[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-15)\n  16. There are in theory several ways the logit weights could overestimate the model's actual use of a feature:   \n1. It could be that these output weights are small enough that, when multiplied by activations, they don't have an appreciable effect on the model’s output.  \n2. The feature might only activate in situations where other features make these tokens extremely unlikely, such that the feature in fact has little effect.   \n3. It is possible that our approximation of linearizing the layer norm (see Framework ) is poor.  \nBased on the subsequent analysis, which confirms the logit weight effects, we do not believe these issues arise in practice.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-16)\n  17. Of course, some features might genuinely be neuron aligned. But we'd like to know that at least some of the features dictionary learning discovers were not trivial.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-17)\n  18. We also tried looking at the neurons which have the largest contribution to the dictionary vector for the feature. However, we found looking at the most correlated neuron to be more reliable – in earlier experiments, we found rare cases where the correlated method found a seemingly similar neuron, while the dictionary method did not. This may be because neurons can have different scales of activations.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-18)\n  19. This makes sense as a superposition strategy: since languages are essentially mutually exclusive, they're natural to put in superposition with each other[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-19)\n  20. By analogy, this also applies to [B/1/1334](https://transformer-circuits.pub/2023/monosemantic-features/vis/b1.html#feature-1334).[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-20)\n  21. \"Activation correlation\" is defined as the feature whose activations across 40,960,000 tokens has the highest Pearson correlation with those of [A/1/3450](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3450).[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-21)\n  22. In order to sample uniformly across the spectrum of feature activations, we divide the activation spectrum into 11 \"activation intervals\" evenly spaced between 0 activation and the maximum activation. We sample uniformly from these intervals.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-22)\n  23. It's worth explicitly stating that our automated interpretability setup was designed to ensure that there's no leak of information about activation patterns, except for the explanation. For example, when predicting new activations, the model cannot see any true activations of that feature.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-23)\n  24. This is distinct from the evaluation strategy of Bills et al., who calculated their correlation of predicted and true activations on a mixture of maximal dataset examples and random samples. For sparse features, which don't fire on most random samples, this effectively tests the model's ability to distinguish a feature's large activations from zero.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-24)\n  25. In instances where Claude predicts a constant score, most often all 0s, a correlation can't be computed and we assign a score of zero which explains the uptick there.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-25)\n  26. with respect to the effect it has on the model outputs see e.g. the activation expected value plot in [Arabic Feature's Activations Specificity Analysis](https://transformer-circuits.pub/2023/monosemantic-features/#feature-arabic-activations).[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-26)\n  27. Naively, simulating a layer with 100k features would be 100,000 times more expensive than sampling a large language model such as Claude 2 or GPT-4 (we'd need to sample the explaining large model once for every feature at every token). At current prices, this would suggest simulating 100k features on a single 4096 token context would cost $12,500–$25,000, and one would presumably need to evaluate over many contexts.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-27)\n  28. We generate a model with random weights by randomly shuffling the entries of each weight matrix of the trained transformer used in Run A. This guarantees that the distributions of individual weights match, and differences are due to structure.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-28)\n  29. From a purely theoretical lens, attention heads can largely implement \"three point functions\" (with two inputs, and an output). MLP layers are well positioned to instead implement N-token conjunctions, perhaps the most extreme of which are context features or token-in-context features. Thus, it is perhaps natural that we see many of these.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-29)\n  30. Token-in-context features may offer a significant opportunity for simplifying analysis of the model – as Elhage et al. note, it may be possible to understand these features as two-dimensional family of features parameterized by a context and a token.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-30)\n  31. Toy Models considered correlated features (see in particular [organization of correlated features](https://transformer-circuits.pub/2022/toy_model/index.html#geometry-organization) and [collapsing of correlated features](https://transformer-circuits.pub/2022/toy_model/index.html#geometry-collapsing)), but only features which were correlated in whether they were active (and not their value if active), and had nothing analogous to the similar \"output actions\" described here. Nonetheless, Toy Models' experiments may be a useful intuition pump, especially in noticing the distinction between similar features in superposition vs features collapsing into a single broader feature.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-31)\n  32. For each pair of features, we compute the cosine similarity between their activations on the subset of tokens for which one of them fires. We repeat this, restricting to the subset on which the other fires, and take the greater of the two. We draw a connection between the features if this measure exceeds ~0.4. This threshold was chosen to balance producing a small enough graph to visualize while also showing some of the richness of feature splitting. We omit the ultralow density features from this analysis.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-32)\n  33. The features fire on the token \" P\" of words where it appears as the first token, such as `[ P][attern]`.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-33)\n  34. In this instance we found the refined `P` features by manual inspection rather than by cosine similarity.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-34)\n  35. Our initial clue that [A/1/1544](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1544) might fire on base64 strings encoding ASCII text was the token `ICAgICAg` which this feature particularly responds to, and corresponds to six spaces in a row.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-35)\n  36. Determining when a dataset example encodes ASCII text is somewhat subtle because base64 can only be decoded to ASCII in groups of four characters, since four base64 characters [encode triples](https://en.wikipedia.org/wiki/Base64#Examples) of ASCII characters. Thus, we select substrings which – when decoded with the python base64 library – contain the maximal number of printable ASCII characters.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-36)\n  37. In what sense does universality suggest features are \"real\"? One basic observation is that they suggest the features we're finding are not just artifacts of the dictionary learning process – or at least that if they come from the dictionary learning process, it's in some consistent way. But there are also several deeper ways in which it's suggestive. It means that, whatever the source of the features, we can talk about features as replicable, reliable, recurring units of analysis. It's also just a surprising observation that one would expect if a strong version of the features in superposition hypothesis was true and models were literally representing some finite, discrete set of features in superposition.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-37)\n  38. In fact, we found some features so universal that we began to take it for granted as a basic tool in our workflow of evaluating dictionary learning runs. For example, the base64 feature – which we previously observed in SoLU models – was so consistently universal that its presence was a useful debugging heuristic.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-38)\n  39. The feature is bimodal, monosemantic in the larger of the modes, and fires on 0.02% of tokens. This means that at least 1 in 10,000 tokens in the Pile dataset are abbreviations for PLoS journals in citations! This is an example of how inspecting features can reveal properties of the dataset, in this case the strong bias of the Pile towards scientific content.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-39)\n  40. Suppose feature f_i has logit weights v_{ik} for k \\in \\\\{1,\\ldots,n_{\\text{vocab}}\\\\}. At a given token t_j, we compute the activation of the feature f_i(t_j) and multiply it by the logit weight v_{it_{j+1}} of the token t_{j+1} that comes next to get an attribution score of f_i(t_j)v_{it_{j+1}}. The attribution vector is given by stacking the attribution scores for a random sampling of datapoints. This approximates the classic attribution method of multiplying the gradient by the activation, differing in that we ignore the denominators of the softmax and the layer norm.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-40)\n  41. The \"finite state automata\"-esque feature assemblies are also different from circuits in that the model didn't learn them to work together. Rather, in the course of learning to autoregressively model text, it learned features that interact via the token stream because of patterns in the real datasets. In contrast, a language model trained with reinforcement learning might have systems like this – circuits whose feature components interact via generated tokens – which co-evolved and adapted to work together during RL training.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-41)\n  42. The right mode in this example may actually be split into two modes, one around 10^{-5} and another around 10^{-3}. We are still investigating the nature of this split.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-42)\n  43. We are explicit in keeping all of these examples separate from the ones used when we test on activation predictions.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-43)\n  44. The top activating tokens out of context examples inherit the weighting of the top activating examples quantile.[[↩]](https://transformer-circuits.pub/2023/monosemantic-features/#d-footnote-44)\n\n\n### References\n  1. Feature Visualization [[link]](https://distill.pub/2017/feature-visualization)  \nOlah, C., Mordvintsev, A. and Schubert, L., 2017. Distill. [DOI: 10.23915/distill.00007](https://doi.org/10.23915/distill.00007)\n  2. Linear algebraic structure of word senses, with applications to polysemy   \nArora, S., Li, Y., Liang, Y., Ma, T. and Risteski, A., 2018. Transactions of the Association for Computational Linguistics, Vol 6, pp. 483–495. MIT Press.\n  3. Decoding The Thought Vector [[link]](https://gabgoh.github.io/ThoughtVectors/)  \nGoh, G., 2016. \n  4. Zoom In: An Introduction to Circuits   \nOlah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S., 2020. Distill. [DOI: 10.23915/distill.00024.001](https://doi.org/10.23915/distill.00024.001)\n  5. Toy Models of Superposition   \nElhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Hatfield-Dodds, Z., Lasenby, R., Drain, D., Chen, C., Grosse, R., McCandlish, S., Kaplan, J., Amodei, D., Wattenberg, M. and Olah, C., 2022. Transformer Circuits Thread. \n  6. Sparse overcomplete word vector representations   \nFaruqui, M., Tsvetkov, Y., Yogatama, D., Dyer, C. and Smith, N., 2015. arXiv preprint arXiv:1506.02004. \n  7. Spine: Sparse interpretable neural embeddings   \nSubramanian, A., Pruthi, D., Jhamtani, H., Berg-Kirkpatrick, T. and Hovy, E., 2018. Proceedings of the AAAI Conference on Artificial Intelligence, Vol 32(1). \n  8. Word embedding visualization via dictionary learning   \nZhang, J., Chen, Y., Cheung, B. and Olshausen, B.A., 2019. arXiv preprint arXiv:1910.03833. \n  9. Word2Sense: sparse interpretable word embeddings   \nPanigrahi, A., Simhadri, H.V. and Bhattacharyya, C., 2019. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 5692–5705. \n  10. Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors   \nYun, Z., Chen, Y., Olshausen, B.A. and LeCun, Y., 2021. arXiv preprint arXiv:2103.15949. \n  11. [Interim research report] Taking features out of superposition with sparse autoencoders [[link]](https://www.lesswrong.com/posts/z6QQJbtpkEAX3Aojj/interim-research-report-taking-features-out-of-superposition)  \nSharkey, L., Braun, D. and Millidge, B., 2022. \n  12. [Replication] Conjecture's Sparse Coding in Toy Models [[link]](https://www.lesswrong.com/posts/E8imxQo96WgDCMxkA/replication-conjecture-s-sparse-coding-in-toy-models)  \nCunningham, H. and Smith, L., 2023. \n  13. [Research Update] Sparse Autoencoder features are bimodal [[link]](https://aizi.substack.com/p/research-update-sparse-autoencoder)  \nHuben, R., 2023. \n  14. (tentatively) Found 600+ Monosemantic Features in a Small LM Using Sparse Autoencoders [[link]](https://www.lesswrong.com/posts/wqRqb7h6ZC48iDgfK/tentatively-found-600-monosemantic-features-in-a-small-lm)  \nSmith, L., 2023. \n  15. Really Strong Features Found in Residual Stream [[link]](https://www.lesswrong.com/posts/Q76CpqHeEMykKpFdB/really-strong-features-found-in-residual-stream)  \nSmith, L., 2023. \n  16. AutoInterpretation Finds Sparse Coding Beats Alternatives [[link]](https://www.lesswrong.com/posts/ursraZGcpfMjCXtnn/autointerpretation-finds-sparse-coding-beats-alternatives)  \nCunningham, H., 2023. \n  17. Sparse Autoencoders Find Highly Interpretable Model Directions [[PDF]](https://arxiv.org/pdf/2309.08600.pdf)  \nCunningham, H., Ewart, A., Smith, L., Huben, R. and Sharkey, L., 2023. arXiv preprint arXiv:2309.08600. \n  18. A Mathematical Framework for Transformer Circuits [[HTML]](https://transformer-circuits.pub/2021/framework/index.html)  \nElhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., DasSarma, N., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Clark, J., Kaplan, J., McCandlish, S. and Olah, C., 2021. Transformer Circuits Thread. \n  19. The Pile: An 800GB Dataset of Diverse Text for Language Modeling   \nGao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., Presser, S. and Leahy, C., 2020. \n  20. Linguistic regularities in continuous space word representations [[PDF]](https://aclanthology.org/N13-1090.pdf)  \nMikolov, T., Yih, W. and Zweig, G., 2013. Proceedings of the 2013 conference of the north american chapter of the association for computational linguistics: Human language technologies, pp. 746–751. \n  21. Linguistic regularities in sparse and explicit word representations   \nLevy, O. and Goldberg, Y., 2014. Proceedings of the eighteenth conference on computational natural language learning, pp. 171–180. \n  22. Unsupervised representation learning with deep convolutional generative adversarial networks   \nRadford, A., Metz, L. and Chintala, S., 2015. arXiv preprint arXiv:1511.06434. \n  23. Visualizing and understanding recurrent networks [[PDF]](https://arxiv.org/pdf/1506.02078.pdf)  \nKarpathy, A., Johnson, J. and Fei-Fei, L., 2015. arXiv preprint arXiv:1506.02078. \n  24. Learning to generate reviews and discovering sentiment [[PDF]](https://arxiv.org/pdf/1704.01444.pdf)  \nRadford, A., Jozefowicz, R. and Sutskever, I., 2017. arXiv preprint arXiv:1704.01444. \n  25. Curve Detectors [[link]](https://distill.pub/2020/circuits/curve-detectors)  \nCammarata, N., Goh, G., Carter, S., Schubert, L., Petrov, M. and Olah, C., 2020. Distill. \n  26. Object detectors emerge in deep scene cnns [[PDF]](https://arxiv.org/pdf/1412.6856.pdf)  \nZhou, B., Khosla, A., Lapedriza, A., Oliva, A. and Torralba, A., 2014. arXiv preprint arXiv:1412.6856. \n  27. Network Dissection: Quantifying Interpretability of Deep Visual Representations [[PDF]](https://arxiv.org/pdf/1704.05796.pdf)  \nBau, D., Zhou, B., Khosla, A., Oliva, A. and Torralba, A., 2017. Computer Vision and Pattern Recognition. \n  28. Understanding the role of individual units in a deep neural network   \nBau, D., Zhu, J., Strobelt, H., Lapedriza, A., Zhou, B. and Torralba, A., 2020. Proceedings of the National Academy of Sciences, Vol 117(48), pp. 30071–30078. National Acad Sciences.\n  29. On the importance of single directions for generalization [[PDF]](https://arxiv.org/pdf/1803.06959.pdf)  \nMorcos, A.S., Barrett, D.G., Rabinowitz, N.C. and Botvinick, M., 2018. arXiv preprint arXiv:1803.06959. \n  30. On Interpretability and Feature Representations: An Analysis of the Sentiment Neuron   \nDonnelly, J. and Roegiest, A., 2019. European Conference on Information Retrieval, pp. 795–802. \n  31. Emergent Linear Representations in World Models of Self-Supervised Sequence Models   \nNanda, N., Lee, A. and Wattenberg, M., 2023. arXiv preprint arXiv:2309.00941. \n  32. Discovering latent knowledge in language models without supervision   \nBurns, C., Ye, H., Klein, D. and Steinhardt, J., 2022. arXiv preprint arXiv:2212.03827. \n  33. Acquisition of chess knowledge in alphazero   \nMcGrath, T., Kapishnikov, A., Toma{\\v{s}}ev, N., Pearce, A., Wattenberg, M., Hassabis, D., Kim, B., Paquet, U. and Kramnik, V., 2022. Proceedings of the National Academy of Sciences, Vol 119(47), pp. e2206625119. National Acad Sciences.\n  34. Emergent world representations: Exploring a sequence model trained on a synthetic task   \nLi, K., Hopkins, A.K., Bau, D., Viégas, F., Pfister, H. and Wattenberg, M., 2022. arXiv preprint arXiv:2210.13382. \n  35. Engineering monosemanticity in toy models   \nJermyn, A.S., Schiefer, N. and Hubinger, E., 2022. arXiv preprint arXiv:2211.09169. \n  36. Polysemanticity and capacity in neural networks   \nScherlis, A., Sachan, K., Jermyn, A.S., Benton, J. and Shlegeris, B., 2022. arXiv preprint arXiv:2210.01892. \n  37. Sparse coding with an overcomplete basis set: A strategy employed by V1?   \nOlshausen, B.A. and Field, D.J., 1997. Vision research, Vol 37(23), pp. 3311–3325. Elsevier.\n  38. Softmax Linear Units   \nElhage, N., Hume, T., Olsson, C., Nanda, N., Henighan, T., Johnston, S., ElShowk, S., Joseph, N., DasSarma, N., Mann, B., Hernandez, D., Askell, A., Ndousse, K., Jones, A., Drain, D., Chen, A., Bai, Y., Ganguli, D., Lovitt, L., Hatfield-Dodds, Z., Kernion, J., Conerly, T., Kravec, S., Fort, S., Kadavath, S., Jacobson, J., Tran-Johnson, E., Kaplan, J., Clark, J., Brown, T., McCandlish, S., Amodei, D. and Olah, C., 2022. Transformer Circuits Thread. \n  39. Sparse and redundant representations: from theory to applications in signal and image processing   \nElad, M., 2010. Vol 2(1). Springer.\n  40. Method of optimal directions for frame design   \nEngan, K., Aase, S.O. and Husoy, J.H., 1999. 1999 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings. ICASSP99 (Cat. No. 99CH36258), Vol 5, pp. 2443–2446. \n  41. K-SVD: An algorithm for designing overcomplete dictionaries for sparse representation   \nAharon, M., Elad, M. and Bruckstein, A., 2006. IEEE Transactions on signal processing, Vol 54(11), pp. 4311–4322. IEEE.\n  42. Attribution Patching: Activation Patching At Industrial Scale [[link]](https://www.neelnanda.io/mechanistic-interpretability/attribution-patching)  \nNanda, N., 2023. \n  43. SolidGoldMagikarp (plus, prompt generation) [[link]](https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldmagikarp-plus-prompt-generation)  \nRumbelow, J. and Watkins, M., 2023. \n  44. Distributed Representations: Composition & Superposition [[HTML]](https://transformer-circuits.pub/2023/superposition-composition/index.html)  \nOlah, C., 2023. \n  45. Language models can explain neurons in language models [[HTML]](https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html)  \nBills, S., Cammarata, N., Mossing, D., Tillman, H., Gao, L., Goh, G., Sutskever, I., Leike, J., Wu, J. and Saunders, W., 2023. \n  46. Natural language descriptions of deep visual features   \nHernandez, E., Schwettmann, S., Bau, D., Bagashvili, T., Torralba, A. and Andreas, J., 2021. International Conference on Learning Representations. \n  47. An Overview of Early Vision in InceptionV1   \nOlah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S., 2020. Distill. [DOI: 10.23915/distill.00024.002](https://doi.org/10.23915/distill.00024.002)\n  48. Multimodal Neurons in Artificial Neural Networks   \nGoh, G., Cammarata, N., Voss, C., Carter, S., Petrov, M., Schubert, L., Radford, A. and Olah, C., 2021. Distill. [DOI: 10.23915/distill.00030](https://doi.org/10.23915/distill.00030)\n  49. Finding Neurons in a Haystack: Case Studies with Sparse Probing   \nGurnee, W., Nanda, N., Pauly, M., Harvey, K., Troitskii, D. and Bertsimas, D., 2023. arXiv preprint arXiv:2305.01610. \n  50. Visualizing and measuring the geometry of BERT   \nCoenen, A., Reif, E., Yuan, A., Kim, B., Pearce, A., Viégas, F. and Wattenberg, M., 2019. Advances in Neural Information Processing Systems, Vol 32. \n  51. Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space   \nGeva, M., Caciularu, A., Wang, K.R. and Goldberg, Y., 2022. arXiv preprint arXiv:2203.14680. \n  52. High-Low Frequency Detectors   \nSchubert, L., Voss, C., Cammarata, N., Goh, G. and Olah, C., 2021. Distill. [DOI: 10.23915/distill.00024.005](https://doi.org/10.23915/distill.00024.005)\n  53. Convergent learning: Do different neural networks learn the same representations?   \nLi, Y., Yosinski, J., Clune, J., Lipson, H., Hopcroft, J.E. and others, 2015. FE@ NIPS, pp. 196–212. \n  54. Why does unsupervised pre-training help deep learning? [[link]](https://www.jmlr.org/papers/volume11/erhan10a/erhan10a)  \nErhan, D., Courville, A., Bengio, Y. and Vincent, P., 2010. Proceedings of the thirteenth international conference on artificial intelligence and statistics, pp. 201–208. \n  55. Visualizing Representations: Deep Learning and Human Beings [[link]](http://colah.github.io/posts/2015-01-Visualizing-Representations/)  \nOlah, C., 2015. \n  56. SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability [[PDF]](http://papers.nips.cc/paper/7188-svcca-singular-vector-canonical-correlation-analysis-for-deep-learning-dynamics-and-interpretability.pdf)  \nRaghu, M., Gilmer, J., Yosinski, J. and Sohl-Dickstein, J., 2017. Advances in Neural Information Processing Systems 30, pp. 6078–6087. Curran Associates, Inc.\n  57. Superposition, Memorization, and Double Descent   \nHenighan, T.A.C., 2023. Transformer Circuits Thread. \n  58. Mechanistic anomaly detection and ELK [[link]](https://ai-alignment.com/mechanistic-anomaly-detection-and-elk-fb84f4c6d0dc)  \nChristiano, P., 2022. \n  59. Dropout can create a privileged basis in the ReLU output model [[link]](https://www.lesswrong.com/posts/uSdFFTATPFJz4pQyB/dropout-can-create-a-privileged-basis-in-the-relu-output)  \nlsgos, 2023. \n  60. More findings on Memorization and double descent [[link]](https://www.alignmentforum.org/posts/KzwB4ovzrZ8DYWgpw/more-findings-on-memorization-and-double-descent)  \nHobbhahn, M., 2023. \n  61. Disentangling by factorising   \nKim, H. and Mnih, A., 2018. International Conference on Machine Learning, pp. 2649–2658. \n  62. Infogan: Interpretable representation learning by information maximizing generative adversarial nets   \nChen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I. and Abbeel, P., 2016. Advances in neural information processing systems, Vol 29. \n  63. k-Sparse Autoencoders [[link]](https://api.semanticscholar.org/CorpusID:14850799)  \nMakhzani, A. and Frey, B.J., 2013. CoRR, Vol abs/1312.5663. \n  64. Representation learning: A review and new perspectives   \nBengio, Y., Courville, A. and Vincent, P., 2013. IEEE transactions on pattern analysis and machine intelligence, Vol 35(8), pp. 1798–1828. IEEE.\n  65. Toward transparent ai: A survey on interpreting the inner structures of deep neural networks   \nRäuker, T., Ho, A., Casper, S. and Hadfield-Menell, D., 2023. 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pp. 464–483. \n  66. Promises and pitfalls of black-box concept learning models   \nMahinpei, A., Clark, J., Lage, I., Doshi-Velez, F. and Pan, W., 2021. arXiv preprint arXiv:2106.13314. \n  67. Local vs. Distributed Coding   \nThorpe, S.J., 1989. Intellectica, Vol 8, pp. 3–40. \n  68. Scaling laws for neural language models   \nKaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J. and Amodei, D., 2020. arXiv preprint arXiv:2001.08361. \n  69. Sparse-Coding Variational Auto-Encoders   \nBarello, G., Charles, A.S. and Pillow, J.W., 2018. bioRxiv. \n  70. On incorporating inductive biases into VAEs   \nMiao, N., Mathieu, E., Siddharth, N., Teh, Y.W. and Rainforth, T., 2021. arXiv preprint arXiv:2106.13746. \n  71. Beyond ℓ₁ sparse coding in V1   \nRentzeperis, I., Calatroni, L., Perrinet, L. and Prandi, D., 2023. arXiv preprint arXiv:2301.10002. \n  72. Emergence of Sparse Representations from Noise   \nBricken, T., Schaeffer, R., Olshausen, B. and Kreiman, G., 2023. \n  73. Adam: A method for stochastic optimization   \nKingma, D.P. and Ba, J., 2014. arXiv preprint arXiv:1412.6980. \n  74. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification   \nHe, K., Zhang, X., Ren, S. and Sun, J., 2015. Proceedings of the IEEE international conference on computer vision, pp. 1026–1034. \n  75. The lottery ticket hypothesis: Finding sparse, trainable neural networks   \nFrankle, J. and Carbin, M., 2018. arXiv preprint arXiv:1803.03635. \n\n\n",
  "body_base_path": "attachments/37d897fcb626/pages/",
  "body_markdown_path": "attachments/37d897fcb626/pages/article.md",
  "local_html": "attachments/37d897fcb626/pages/article.html",
  "body_source": "prior_capture",
  "papers": [
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/1601.03764",
      "bytes": 180478,
      "reused": false,
      "content_type": "application/pdf",
      "local_path": "papers/repair/d977aedebe7a11d3a108.pdf"
    }
  ],
  "body_chars": 162622,
  "expected_body_images": 0,
  "raw_html": "attachments/37d897fcb626/pages/source.html"
}