保留正文与补采附件。源页面当前未重新核验；具体限制见交付清单。
Using a sparse autoencoder, we extract a large number of interpretable features from a one-layer transformer.
Browse A/1 Features →
Browse All Features →
Trenton 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
Oct 4, 2023
Not published yet.
No DOI yet.
Detailed Investigations of Individual Features
Related Work Discussion Comments & Replications
Author Contributions Statement
Acknowledgments
Citation Information
Feature Proxies
Feature Interpretability Rubric
Feature Density Histograms
Transformer Training and Preprocessing
Advice for Training Sparse Autoencoders
Automated Interpretability
Mechanistic 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.
Unfortunately, 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 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.
One 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 , 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.
In Toy Models of Superposition, we described three strategies 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.
In 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. .
The 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.
This report has four major sections. In Problem Setup, we provide motivation for our approach and describe the transformers and sparse autoencoders we train. In Detailed Investigations of Individual Features, 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, we argue that the typical feature is interpretable and that they explain a non-trivial portion of the MLP layer. Finally, in Phenomenology we describe several properties of our features, including feature-splitting, universality, and how they can form "finite state automata"-like systems implementing interesting behaviors.
We also provide three comprehensive visualizations of features. First, for all features from 90 learned dictionaries we present activating dataset examples and downstream logit effects. We recommend the reader begin with the visualization of A/1. Second, we provide a data-oriented view, 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:
Cluster
Feature
Cluster #49
A/0/307
This 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.
Zoom to ↗ View details
A/0/311
This feature fires for reference citations in academic papers, specifically when it sees the [@ symbol that is commonly used.
Zoom to ↗ View details
A/1/776
Years in some citation notation
Zoom to ↗ View details
A/1/1538
Citations in a [@author] or [@authoryear] format
Zoom to ↗ View details
A/1/1875
Markdown Citation (Predict year)
Zoom to ↗ View details
A/1/2252
" [@"
Zoom to ↗ View details
A/1/2237
[Ultralow density cluster]
Zoom to ↗ View details
Cluster #42
A/0/126
This feature seems to fire on section headings, specifically the word "sec" within Markdown section tags like {#sec:foo}.
Zoom to ↗ View details
A/1/357
"ref" in [context]
Zoom to ↗ View details
A/1/1469
"s"/"sec" after "{#", section reference in some markup
Zoom to ↗ View details
A/1/3841
"Sec"
Zoom to ↗ View details
A/1/3898
Section number in {#SecX}
Zoom to ↗ View details
A/1/4083
" {#"
Zoom to ↗ View details
A/1/2129
"." in [context]
Zoom to ↗ View details
A/1/553
"](#" in [context]
Zoom to ↗ View details
Cluster #43
A/0/8
This 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.
Zoom to ↗ View details
A/0/398
This feature attends to references to figures and tables.
Zoom to ↗ View details
A/0/454
This 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.
Zoom to ↗ View details
A/1/35
"){"
Zoom to ↗ View details
A/1/366
"type"
Zoom to ↗ View details
A/1/945
"ref" in [context]
Zoom to ↗ View details
A/1/1895
"-" in [context]
Zoom to ↗ View details
A/1/2176
"fig"
Zoom to ↗ View details
Cluster #46
A/0/425
This feature attends to highlighting bibliography and reference markers in the text such as [1], (Table 1), [@B23], etc.
Zoom to ↗ View details
A/1/703
This feature attends to bibliographic citations with brackets and @CIT numbers.
Zoom to ↗ View details
A/1/864
" \\[[@"
Zoom to ↗ View details
A/1/868
"[@" in bibliography citations
Zoom to ↗ View details
A/1/2217
".]("
Zoom to ↗ View details
A/1/2777
" ([@" in [context]
Zoom to ↗ View details
A/1/3120
This 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.
Zoom to ↗ View details
A/1/3184
Introducing an acaemic reference format
Zoom to ↗ View details
A/1/3698
"](#" in [context]
Zoom to ↗ View details
A/1/51
[Ultralow density cluster]
Zoom to ↗ View details
Cluster #48
A/0/363
The feature attends to in-text citations with URL and DOI links.
Zoom to ↗ View details
A/0/366
This feature fires on citations and reference markers such as [ ], ( ), etc.
Zoom to ↗ View details
A/0/396
This feature fires for in-text references, citation links, and reference brackets/notations in scholarly articles, research papers, and academic journals.
Zoom to ↗ View details
A/1/90
"1"
Zoom to ↗ View details
A/1/204
"CR"
Zoom to ↗ View details
A/1/220
This feature attends to formats of references and citations.
Zoom to ↗ View details
A/1/729
Digits in bibliographies?
Zoom to ↗ View details
A/1/1015
Reference IDs in parentheses in {} citation format
Zoom to ↗ View details
A/1/1264
Mostly letter fragments of IDs in parentheses in {} citation format
Zoom to ↗ View details
A/1/1316
A specific form of medical journal document reference, syntax '{...', specifically the journalid12345-1234 part
Zoom to ↗ View details
A/1/1440
"g" in [context]
Zoom to ↗ View details
A/1/1445
"ijms"
Zoom to ↗ View details
A/1/2200
The feature fires on citation indices in scientific papers, such as [B13-ijms-20-04353-f001].
Zoom to ↗ View details
A/1/2280
"bib" in [context]
Zoom to ↗ View details
A/1/2385
"bib" in some bibliography format ending in "bib-XXXX" where X are digits
Zoom to ↗ View details
A/1/2496
"B" in [context]
Zoom to ↗ View details
A/1/2559
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2696
Numbers/digits in hyphenated IDs in square brackets, likely [@ ... ] citation format
Zoom to ↗ View details
A/1/2881
"B" in [context]
Zoom to ↗ View details
A/1/2988
Digits after upper case letters
Zoom to ↗ View details
A/1/2996
This feature attends to inline citations like [R9] or [@ref9] in scientific papers.
Zoom to ↗ View details
A/1/3143
This feature seems to fire on citations and references in brackets and parentheses with numbers, IDs and abbreviations.
Zoom to ↗ View details
A/1/3163
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3164
Numbers after "pone.", for citation IDs
Zoom to ↗ View details
A/1/3547
Beginnings of hash locations in Markdown URL targets?
Zoom to ↗ View details
A/1/3949
"pone"
Zoom to ↗ View details
A/1/4062
Post-"#" numeric ID fragments in parentheses in {} citation format
Zoom to ↗ View details
A/0/270
This feature fires on reference links to articles, papers, and datasets in academic writing. These references links are typically in a specific format like text or with abbreviated journal titles and citation numbers like [B14-sensors-19-00441].
Zoom to ↗ View details
Cluster #36
A/0/277
The feature fires when it sees short HTML/XML tags such as <ref-type="fig">
Zoom to ↗ View details
A/0/427
This feature appears to fire on hyphens between numbers, especially years.
Zoom to ↗ View details
A/1/367
"="" in [context]
Zoom to ↗ View details
A/1/747
"-" in [context]
Zoom to ↗ View details
A/1/1220
"-" in [context]
Zoom to ↗ View details
A/1/1782
"-" in [context]
Zoom to ↗ View details
A/1/2324
"="" in [context]
Zoom to ↗ View details
A/1/2876
"-" in [context]
Zoom to ↗ View details
A/1/3122
"-" for dashes within journal citations?
Zoom to ↗ View details
A/1/774
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3838
" @" in [context]
Zoom to ↗ View details
A/1/1415
"),", sometimes with more leading punctuation, in scientific contexts
Zoom to ↗ View details
A/1/3877
"," in [context]
Zoom to ↗ View details
Cluster #124
A/1/1147
"," in [context]
Zoom to ↗ View details
A/1/2930
"," in [context]
Zoom to ↗ View details
A/1/3857
"," in [context]
Zoom to ↗ View details
Cluster #127
A/0/359
This feature fires when it sees the punctuation mark comma ","
Zoom to ↗ View details
A/1/178
"," in mechanical engineering?
Zoom to ↗ View details
A/1/713
"," after "However" or "by contrast" type phrases
Zoom to ↗ View details
A/1/1204
This feature attends to commas used as enumerators in scientific texts.
Zoom to ↗ View details
A/1/1374
"," in lists of chemicals
Zoom to ↗ View details
A/1/1565
"],"
Zoom to ↗ View details
A/1/1801
"," between demographic factors
Zoom to ↗ View details
A/1/2284
"," in [context]
Zoom to ↗ View details
A/1/2835
"," in [context]
Zoom to ↗ View details
A/1/3482
";" in [context]
Zoom to ↗ View details
A/1/3921
";" in [context]
Zoom to ↗ View details
Cluster #37
A/0/289
This feature attends to parentheses and other bracket notations.
Zoom to ↗ View details
A/1/1785
This feature attends to closing parentheses.
Zoom to ↗ View details
A/1/1885
")" in biomedical studies
Zoom to ↗ View details
A/1/3672
. The "]" at the end of each reference code
Zoom to ↗ View details
A/1/715
"," after an adverb indicating a conclusion from something previous, e.g. "Finally"
Zoom to ↗ View details
A/0/231
The feature attends to the word "in" followed by a comma and appearing at the beginning of a sentence.
Zoom to ↗ View details
Cluster #142
A/0/430
This feature fires when a comma is present.
Zoom to ↗ View details
A/1/1284
"," in [context]
Zoom to ↗ View details
A/1/1479
"," in [context]
Zoom to ↗ View details
A/1/2409
"," in [context]
Zoom to ↗ View details
A/1/3342
"," in lists, mostly of short nouns/phrases
Zoom to ↗ View details
A/1/3411
"," in [context]
Zoom to ↗ View details
Cluster #10
A/0/128
This feature appears to attend to proper nouns and named entities in text, especially names of people, organizations, locations, and creative works.
Zoom to ↗ View details
A/1/626
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1366
Sports, esp. Title Case names of competitions
Zoom to ↗ View details
A/1/1860
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2731
" the" in sports (football?)
Zoom to ↗ View details
A/1/2784
American sports concepts: offensive/defensive lines and defensive/tackle, right-handed bat, three-pointer, starter, drills
Zoom to ↗ View details
A/1/3316
Professional sports
Zoom to ↗ View details
A/1/3336
1- to 2-digit numbers in sports contexts, esp. scores
Zoom to ↗ View details
A/1/3856
" with" in sports contexts
Zoom to ↗ View details
A/1/993
"," in [context]
Zoom to ↗ View details
A/1/1814
"," in [context]
Zoom to ↗ View details
A/1/757
Commas after conjunctions at starts of sentences or connecting clauses
Zoom to ↗ View details
Cluster #143
A/1/586
Commas in discussions of military battles
Zoom to ↗ View details
A/1/658
This feature fires when there is a comma used in the middle or end of a sentence.
Zoom to ↗ View details
A/1/1327
This feature appears to fire when there is a comma followed by a word starting with a vowel.
Zoom to ↗ View details
A/1/1389
Comma after an "if" clause
Zoom to ↗ View details
A/1/3754
"," in [context]
Zoom to ↗ View details
A/1/3160
"," in [context]
Zoom to ↗ View details
Cluster #6
A/0/75
This feature fires when it encounters the word "and" or the word "or" being used as conjunctions.
Zoom to ↗ View details
A/0/214
The feature fires when it sees the word 'and'.
Zoom to ↗ View details
A/0/403
The feature fires for sentences that have a comma separating two phrases, clauses or lists, or the words "or", "and", "but".
Zoom to ↗ View details
A/1/33
" and" in [context]
Zoom to ↗ View details
A/1/56
Code: " and"/" or" in comments about data
Zoom to ↗ View details
A/1/87
" and" after country names and other geographical locations
Zoom to ↗ View details
A/1/92
" and" in [context]
Zoom to ↗ View details
A/1/310
"and" conjoining family members
Zoom to ↗ View details
A/1/324
" and" after a full (first+last) name
Zoom to ↗ View details
A/1/337
This feature attends to coordinating conjunctions such as "and" and "or".
Zoom to ↗ View details
A/1/391
" and" in U.S. history or geography
Zoom to ↗ View details
A/1/406
" and" in [context]
Zoom to ↗ View details
A/1/524
" or" joining number words or similar, e.g. "one or two" or "one or more"
Zoom to ↗ View details
A/1/603
"and" in chemistry text
Zoom to ↗ View details
A/1/783
[Ultralow density cluster]
Zoom to ↗ View details
A/1/802
" or" in [context]
Zoom to ↗ View details
A/1/822
[Ultralow density cluster]
Zoom to ↗ View details
A/1/843
" and" in recipes
Zoom to ↗ View details
A/1/1000
" and" and other conjunctions in computing text
Zoom to ↗ View details
A/1/1074
" and" in [context]
Zoom to ↗ View details
A/1/1137
" and" shortly after EOT
Zoom to ↗ View details
A/1/1211
" and" in [context]
Zoom to ↗ View details
A/1/1249
" and" scientific
Zoom to ↗ View details
A/1/1274
" and" in [context]
Zoom to ↗ View details
A/1/1349
" and" in [context]
Zoom to ↗ View details
A/1/1386
" and" in [context]
Zoom to ↗ View details
A/1/1431
" or" in [context]
Zoom to ↗ View details
A/1/1495
" and" in [context]
Zoom to ↗ View details
A/1/1851
" and" in [context]
Zoom to ↗ View details
A/1/1882
" and" in [context]
Zoom to ↗ View details
A/1/1934
" and" after an emotion, predicting another emotion
Zoom to ↗ View details
A/1/1949
" and" in [context]
Zoom to ↗ View details
A/1/2058
" &" in [context]
Zoom to ↗ View details
A/1/2117
"and"
Zoom to ↗ View details
A/1/2208
" and"/" or", after a number and time interval (e.g. "six months") and predicting another time interval
Zoom to ↗ View details
A/1/2251
" and" in [context]
Zoom to ↗ View details
A/1/2272
" and" and conjunctions in demography or population surveys
Zoom to ↗ View details
A/1/2356
" and" in [context]
Zoom to ↗ View details
A/1/2419
" and" after groups of people, predicting other groups of people
Zoom to ↗ View details
A/1/2428
" and" in [context]
Zoom to ↗ View details
A/1/2431
" and" in [context]
Zoom to ↗ View details
A/1/2520
" and" in [context]
Zoom to ↗ View details
A/1/2585
" and" in [context]
Zoom to ↗ View details
A/1/2597
" and" in [context]
Zoom to ↗ View details
A/1/2612
" and" in [context]
Zoom to ↗ View details
A/1/2623
" and" in [context]
Zoom to ↗ View details
A/1/2679
" and" in [context]
Zoom to ↗ View details
A/1/2747
"and" in context of color descriptions of objects
Zoom to ↗ View details
A/1/2849
" or" in [context]
Zoom to ↗ View details
A/1/2860
The feature fires when words like "or", "and", "to", or punctuation like "," appear.
Zoom to ↗ View details
A/1/2884
" and" in [context]
Zoom to ↗ View details
A/1/3042
" and" in [context]
Zoom to ↗ View details
A/1/3106
" and" in [context]
Zoom to ↗ View details
A/1/3142
'or' in 'and/or'
Zoom to ↗ View details
A/1/3196
" and" in [context]
Zoom to ↗ View details
A/1/3217
" and" in [context]
Zoom to ↗ View details
A/1/3220
" and"/" or" after an abbreviation, predicting another abbreviation
Zoom to ↗ View details
A/1/3232
" and"/" or" largely in all-lowercase contexts like chat logs
Zoom to ↗ View details
A/1/3244
" and" in [context]
Zoom to ↗ View details
A/1/3454
" and" in [context]
Zoom to ↗ View details
A/1/3486
" and"/" or" after and predicting adjectives like "economic" or "medical"
Zoom to ↗ View details
A/1/3618
" and" in [context]
Zoom to ↗ View details
A/1/3671
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3757
" and" in [context]
Zoom to ↗ View details
A/1/3790
This feature fires for sentences that contain the word "and".
Zoom to ↗ View details
A/1/3875
"and" conjunction predicting a past-tense verb
Zoom to ↗ View details
A/1/3906
" and" in [context]
Zoom to ↗ View details
A/1/3983
" and" and sometimes " or" in medical/anatomical contexts
Zoom to ↗ View details
A/1/2611
"," in [context]
Zoom to ↗ View details
Cluster #147
A/0/310
This feature fires when there is a comma followed by a space.
Zoom to ↗ View details
A/1/2367
"," in [context]
Zoom to ↗ View details
A/1/3376
"," in [context]
Zoom to ↗ View details
A/1/3595
This 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.
Zoom to ↗ View details
A/1/3634
"," in [context]
Zoom to ↗ View details
A/1/3995
"," in [context]
Zoom to ↗ View details
A/1/3506
Commas near beginning of quoted English text
Zoom to ↗ View details
A/1/3966
Adjective describing a famous person (occupation or nationality)
Zoom to ↗ View details
Cluster #146
A/0/504
This feature seems to fire when there is a comma followed by the word "and".
Zoom to ↗ View details
A/1/814
"," in [context]
Zoom to ↗ View details
A/1/1059
"," in [context]
Zoom to ↗ View details
A/1/1116
',' comma in archaic english
Zoom to ↗ View details
A/1/1988
"," in [context]
Zoom to ↗ View details
A/1/2925
Female - she / her / herself ... ?
Zoom to ↗ View details
A/1/2390
";" in [context]
Zoom to ↗ View details
A/1/1928
"," in [context]
Zoom to ↗ View details
A/1/382
"," in sports reporting, particularly after "season" / units of time
Zoom to ↗ View details
A/1/3897
"," in [context]
Zoom to ↗ View details
Cluster #105
A/0/123
This feature fires when a name is followed by a comma.
Zoom to ↗ View details
A/1/739
Comma after a work of art (performance, album, book)
Zoom to ↗ View details
A/1/1081
"," separating full names in lists
Zoom to ↗ View details
A/1/1238
Commas in lists of countries
Zoom to ↗ View details
A/1/1631
"," in [context]
Zoom to ↗ View details
A/1/1828
"," after scientific institutions
Zoom to ↗ View details
A/1/2007
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2397
"," in [context]
Zoom to ↗ View details
A/1/2964
Commas in European history
Zoom to ↗ View details
A/1/3459
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1365
The feature fires when it sees quotation marks with text followed by punctuation such as a comma or period.
Zoom to ↗ View details
A/1/657
"," in [context]
Zoom to ↗ View details
Cluster #84
A/1/331
"," in legal contexts?
Zoom to ↗ View details
A/1/673
"," after/between years
Zoom to ↗ View details
A/1/1864
"," in [context]
Zoom to ↗ View details
Cluster #80
A/0/6
The feature fires when it sees punctuation like commas, semicolons and periods, especially when they follow a word that ends with a lowercase letter.
Zoom to ↗ View details
A/1/1066
"," in [context]
Zoom to ↗ View details
A/1/1152
"," in [context]
Zoom to ↗ View details
A/1/3394
This feature fires when it sees a semicolon (;) followed by a parenthetical citation with a 4 digit date.
Zoom to ↗ View details
A/1/3697
"," in [context]
Zoom to ↗ View details
A/1/1692
"," in [context]
Zoom to ↗ View details
A/1/2893
Commas in thousands-place separators (patent numbers?)
Zoom to ↗ View details
Cluster #57
A/1/1406
".," in [context]
Zoom to ↗ View details
A/1/2622
"," after physicist last names, predicting " Phys"
Zoom to ↗ View details
A/1/2767
Commas in citations
Zoom to ↗ View details
Cluster #73
A/0/240
This feature attends to numbers from 1 to 99,999 preceded or followed by commas.
Zoom to ↗ View details
A/1/32
"," after small numbers in sequences
Zoom to ↗ View details
A/1/2014
Commas separating numbers (with spaces)
Zoom to ↗ View details
A/1/2209
"," in numbers
Zoom to ↗ View details
Cluster #83
A/1/546
Commas, often in book titles or preceding them to delimit the author
Zoom to ↗ View details
A/1/2690
"]" in [context]
Zoom to ↗ View details
A/1/3653
")" in legal documents
Zoom to ↗ View details
A/1/4051
Korean: ?
Zoom to ↗ View details
Cluster #7
A/0/48
This feature fires when it sees Greek text, specifically Ancient Greek.
Zoom to ↗ View details
A/1/572
Greek: ?
Zoom to ↗ View details
A/1/634
Greek: "ι"?
Zoom to ↗ View details
A/1/1586
Khmer, Amharic, (Greek?): ?
Zoom to ↗ View details
A/1/2599
Greek: Capital letters?
Zoom to ↗ View details
A/1/3453
Greek: nu, chi
Zoom to ↗ View details
A/1/3831
Greek: Vowel, especially omicron and upsilon
Zoom to ↗ View details
Cluster #13
A/0/66
This feature fires when it sees foreign characters such as accented characters, Cyrillic and Greek characters.
Zoom to ↗ View details
A/0/79
This feature seems to fire on mathematical fractions, degrees symbols, and other symbols commonly used in math and science writing.
Zoom to ↗ View details
A/0/80
This feature fires on Greek text.
Zoom to ↗ View details
A/0/101
This 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.
Zoom to ↗ View details
A/0/290
This feature attends to non-Latin alphabet languages such as Thai, Tamil, Georgian, Hindi, Japanese, Arabic, etc.
Zoom to ↗ View details
A/1/0
Japanese, Chinese, other Asian scripts: Interwoven in English
Zoom to ↗ View details
A/1/78
Japanese: "\xe9"?
Zoom to ↗ View details
A/1/281
Greek?: "\\xe1"
Zoom to ↗ View details
A/1/422
Thai: Unicode prefix "\xe0\xb8", after vowel
Zoom to ↗ View details
A/1/426
Start of Unicode character esp. predicting \x95 in U+2215 HORIZONTAL BAR
Zoom to ↗ View details
A/1/432
"\xc3" in UTF-8 as latin1 mojibake, esp. of Russian
Zoom to ↗ View details
A/1/558
"\\xce" in [context]
Zoom to ↗ View details
A/1/589
"\\xe0\\xa4"
Zoom to ↗ View details
A/1/619
"\\xe1\\xbb"
Zoom to ↗ View details
A/1/641
Japanese: beginnings of characters?
Zoom to ↗ View details
A/1/795
Arabic: Unicode start "\xd9"
Zoom to ↗ View details
A/1/870
Japanese: ?
Zoom to ↗ View details
A/1/903
"\\xce" in [context]
Zoom to ↗ View details
A/1/934
Japanese: "\xe8\xa6"?
Zoom to ↗ View details
A/1/1009
" \\xce"
Zoom to ↗ View details
A/1/1016
Hebrew, Arabic, Farsi: Unicode block "\xd7"?
Zoom to ↗ View details
A/1/1096
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1163
Telugu?: ?
Zoom to ↗ View details
A/1/1193
Japanese: "\\xe3\\x82"?
Zoom to ↗ View details
A/1/1232
Math (Incomplete Unicode)
Zoom to ↗ View details
A/1/1235
\xc2 in UTF-8 as latin1 mojibake of Korean?
Zoom to ↗ View details
A/1/1345
"\xc3" in UTF-8 as latin1 mojibake of Arabic?
Zoom to ↗ View details
A/1/1369
Bengali: Unicode sequence start
Zoom to ↗ View details
A/1/1499
"\\xc4" in [context]
Zoom to ↗ View details
A/1/1724
Georgian and Korean?: "\xe1\x83"
Zoom to ↗ View details
A/1/1820
South Asian scripts: Unicode start
Zoom to ↗ View details
A/1/1904
Japanese: Katakana "extend the previous vowel sound for longer" partial-uniocde character
Zoom to ↗ View details
A/1/1940
Japanese: beginnings of kanji
Zoom to ↗ View details
A/1/2012
"\\xcf" in [context]
Zoom to ↗ View details
A/1/2152
Japanese: Beginnings of kanji
Zoom to ↗ View details
A/1/2399
Japanese, Chinese: "\xe6\xa5"?
Zoom to ↗ View details
A/1/2453
Japanese: "\xe9\x9a"?
Zoom to ↗ View details
A/1/2465
Armenian: "\xd5"
Zoom to ↗ View details
A/1/2466
\xc2 as UTF-8 as latin1 mojibake of Chinese?
Zoom to ↗ View details
A/1/2472
Vietnamese (Thai?): "\xe1\xba"
Zoom to ↗ View details
A/1/2556
Japanese, Chinese, other Asian scripts: Unicode Start \xe6\xb0?
Zoom to ↗ View details
A/1/2563
Thai: Unicode prefix "\xe0\xb9"
Zoom to ↗ View details
A/1/2608
"\\xce" in [context]
Zoom to ↗ View details
A/1/2617
Chinese and Japanese: beginnings of characters
Zoom to ↗ View details
A/1/2618
Japanese: Hiragana unicode character byte-prefix (predicting the postfix)
Zoom to ↗ View details
A/1/2636
\xc2 in UTF-8 as latin1 mojibake of Korean?
Zoom to ↗ View details
A/1/2725
\xc2 in UTF-8 as latin1 mojibake of Chinese?
Zoom to ↗ View details
A/1/2736
Japanese and Chinese: "\xe5\x88"
Zoom to ↗ View details
A/1/2745
\xc2 in UTF-8 as latin1 mojibake of Chinese?
Zoom to ↗ View details
A/1/2782
Japanese: "\\xe3\\x83"
Zoom to ↗ View details
A/1/2798
\xc2 (Unicode starter of Unicode continuation) in UTF-8 as latin1 mojibake of Russian?
Zoom to ↗ View details
A/1/2852
" \\xcf"
Zoom to ↗ View details
A/1/2919
" \\xe0\\xa4"
Zoom to ↗ View details
A/1/2957
Unicode start codes for combining accents
Zoom to ↗ View details
A/1/3006
Emoji? / mid-Unicode bytes
Zoom to ↗ View details
A/1/3133
Hebrew: "\xd6"?
Zoom to ↗ View details
A/1/3150
"\\xc5"
Zoom to ↗ View details
A/1/3345
Gujarati: ?
Zoom to ↗ View details
A/1/3368
Japanese and other Asian languages: "\xe9\x80"?
Zoom to ↗ View details
A/1/3399
Arabic: Unicode start "\xd8"?
Zoom to ↗ View details
A/1/3468
Japanese, Chinese, and Korean: "\\xe7"?
Zoom to ↗ View details
A/1/3561
Thai: unicode prefix "\xe0\xb8"
Zoom to ↗ View details
A/1/3687
Georgian (Unicode Start)
Zoom to ↗ View details
A/1/3692
Hebrew: Unicode Block " \xd7"?
Zoom to ↗ View details
A/1/3726
"\xc2" in UTF-8 as latin1 mojibake of Chinese?
Zoom to ↗ View details
A/1/3748
" \xcc" in mojibake?
Zoom to ↗ View details
A/1/3785
Japanese: "\xe6\x8a"?
Zoom to ↗ View details
A/1/3815
Chinese and Japanese: Beginnings of characters, vaguely poetic?
Zoom to ↗ View details
A/1/3907
Arabic: Unicode start " \xd8"?
Zoom to ↗ View details
A/1/3908
" \xc3" in UTF-8 as latin1 mojibake of Korean? (beginnings of chat logs?)
Zoom to ↗ View details
A/1/3967
Japanese: "\xe7\x9b"?
Zoom to ↗ View details
A/1/4004
Japanese: "\xe8\xaa"?
Zoom to ↗ View details
A/1/4050
Greek: ?
Zoom to ↗ View details
A/1/522
Tamil: Unicode start
Zoom to ↗ View details
A/1/1280
Greek: ?
Zoom to ↗ View details
A/1/3119
Czech, Lithuanian, Tamil: Unicode code point start
Zoom to ↗ View details
Cluster #17
A/0/435
The feature fires when it sees text written in South Indian languages such as Tamil, Telugu, Malayalam, and Kannada.
Zoom to ↗ View details
A/1/416
Hebrew: Unicode end
Zoom to ↗ View details
A/1/466
Sinhala: ?
Zoom to ↗ View details
A/1/480
Armenian: Unicode end
Zoom to ↗ View details
A/1/580
Georgian: ?
Zoom to ↗ View details
A/1/994
Georgian: ?
Zoom to ↗ View details
A/1/1287
Bengali: Unicode end
Zoom to ↗ View details
A/1/1900
Punjabi: ?
Zoom to ↗ View details
A/1/2499
Tamil: Unicode end
Zoom to ↗ View details
A/1/2586
Hindi: ?
Zoom to ↗ View details
A/1/2629
Bengali, Thai, Tamil, Lao?: ?
Zoom to ↗ View details
A/1/2790
Devanagari (Hindi?): ?
Zoom to ↗ View details
A/1/2796
Devanagari, Tamil, Telugu, Malayalam, and Thai: Unicode end
Zoom to ↗ View details
Cluster #4
A/0/17
The feature fires on non-ascii and special characters.
Zoom to ↗ View details
A/1/873
UTF-8 as latin1 mojibake of Chinese?
Zoom to ↗ View details
A/1/980
Euro sign etc. in mojibake: UTF-8 as latin1 of quotes, U+201x
Zoom to ↗ View details
A/1/1049
UTF-8 as latin1 mojibake of Chinese?
Zoom to ↗ View details
A/1/1411
"ä" U+00E4 in UTF-8 as latin1 mojibake, esp. of Chinese
Zoom to ↗ View details
A/1/1517
"\x91" in UTF-8 as latin1 mojibake, esp. of Russian
Zoom to ↗ View details
A/1/1635
"\\x90" UTF-8 as latin1 mojibake of Russian?
Zoom to ↗ View details
A/1/2387
UTF-8 as latin1 mojibake of Korean?
Zoom to ↗ View details
A/1/3041
Particularly nasty mojibake (latin1, Windows-1252, Mac Roman; possibly iterated multiple times)
Zoom to ↗ View details
A/1/3440
"Ã" U+00C3 in two-character mojibake, broadly accented characters in European languages
Zoom to ↗ View details
A/1/3914
UTF-8 as latin1 mojibake of Chinese?
Zoom to ↗ View details
A/1/3961
Some Cyrillic script? Unicode starters in UTF-8 as latin1 mojibake, extra cursed by normalization
Zoom to ↗ View details
A/1/1691
Korean: " \\xed"/"\\xed"
Zoom to ↗ View details
Cluster #15
A/0/319
The feature attends to Korean text.
Zoom to ↗ View details
A/1/1184
Korean: ?
Zoom to ↗ View details
A/1/2323
Arabic, Korean, Chinese, Japanese, Russian, Norwegian, et al.: ?
Zoom to ↗ View details
A/1/2955
Korean: ?
Zoom to ↗ View details
Cluster #18
A/0/459
The feature appears to activate for text written in right-to-left languages such as Arabic, Persian, and Hebrew.
Zoom to ↗ View details
A/1/1271
Arabic script: Whitespace
Zoom to ↗ View details
A/1/1466
Arabic script (Urdu?): ?
Zoom to ↗ View details
A/1/1706
Thai: (Unicode Sequence Complete)
Zoom to ↗ View details
A/1/3134
Arabic script: ?
Zoom to ↗ View details
A/1/3450
Arabic script: ?
Zoom to ↗ View details
Cluster #20
A/1/207
Turkic family?: languages in Cyrillic script
Zoom to ↗ View details
A/1/1197
Russian: " \u0441"
Zoom to ↗ View details
A/1/1858
Russian: ?
Zoom to ↗ View details
A/1/2363
Russian: in (conjunction and verbal prefix)
Zoom to ↗ View details
A/1/2459
Russian: word-ends?
Zoom to ↗ View details
A/1/766
Russian: ?
Zoom to ↗ View details
Cluster #14
A/0/98
This feature attends to Russian words related to action verbs, especially those ending in -ть.
Zoom to ↗ View details
A/0/413
This feature attends to Russian grammatical features like verb conjugations and declensions.
Zoom to ↗ View details
A/1/3181
Russian: Prefix in verbs / nouns
Zoom to ↗ View details
A/1/3061
Russian: Beginnings of quotes
Zoom to ↗ View details
A/1/2664
Russian: " \u0443"
Zoom to ↗ View details
Cluster #19
A/1/24
Russian: "и" ?
Zoom to ↗ View details
A/1/1473
Russian: " Т" / " т"
Zoom to ↗ View details
A/1/1964
Russian: " \u0437"
Zoom to ↗ View details
A/1/2557
Russian?: The common "-ция" suffix indicates that the word is a noun
Zoom to ↗ View details
A/1/3073
Russian: л and к characters?
Zoom to ↗ View details
A/1/3415
Russian: Mostly letter "е"
Zoom to ↗ View details
A/1/3553
Russian: " \u0434"
Zoom to ↗ View details
A/1/4095
Russian: " п"
Zoom to ↗ View details
A/0/130
This feature fires when it sees contractions with "I'm" or "I am", especially at the beginning of sentences or quotes.
Zoom to ↗ View details
A/1/225
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1002
" am" after "I", in context of subjective experience
Zoom to ↗ View details
A/1/983
" im"
Zoom to ↗ View details
Cluster #31
A/0/55
This feature fires when the words "was" or "were" have a high activation value.
Zoom to ↗ View details
A/0/164
This feature fires when the token "is" is present.
Zoom to ↗ View details
A/0/272
This feature fires when the phrase "be _" (with some variation) is present.
Zoom to ↗ View details
A/1/4
" was", esp. after "he"/"she"/a name, predicting an adjective
Zoom to ↗ View details
A/1/111
" is" in [context]
Zoom to ↗ View details
A/1/253
" be" after auxiliary verbs like "must"/"may", sometimes also with "not"
Zoom to ↗ View details
A/1/517
" is" in [context]
Zoom to ↗ View details
A/1/797
" be" in [context]
Zoom to ↗ View details
A/1/885
" is" in [context]
Zoom to ↗ View details
A/1/1095
" is" in [context]
Zoom to ↗ View details
A/1/1378
The 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.
Zoom to ↗ View details
A/1/1521
" been" in [context]
Zoom to ↗ View details
A/1/1750
" become"/" becoming"
Zoom to ↗ View details
A/1/2049
" is" in [context]
Zoom to ↗ View details
A/1/2076
This feature fires when a verb in the past tense follows "was" or "were".
Zoom to ↗ View details
A/1/2131
" was" in [context]
Zoom to ↗ View details
A/1/2188
" is" in [context]
Zoom to ↗ View details
A/1/2325
" is" in [context]
Zoom to ↗ View details
A/1/2481
"al" - logit nonsense
Zoom to ↗ View details
A/1/2755
" is" in [context]
Zoom to ↗ View details
A/1/3015
" is" in [context]
Zoom to ↗ View details
A/1/3063
" be" in [context]
Zoom to ↗ View details
A/1/3161
" are" in [context]
Zoom to ↗ View details
A/1/3535
" is" / other copulae, esp after <EOT>
Zoom to ↗ View details
A/1/3558
" be" in [context]
Zoom to ↗ View details
A/1/3910
" was" in [context]
Zoom to ↗ View details
A/1/3982
" being"
Zoom to ↗ View details
A/1/2560
" remain"
Zoom to ↗ View details
Cluster #32
A/0/295
This feature fires when the token "was" appears in the text.
Zoom to ↗ View details
A/1/1001
Copulas (mostly " is") in technical text
Zoom to ↗ View details
A/1/1146
" is" in [context]
Zoom to ↗ View details
A/1/1426
" is" in [context]
Zoom to ↗ View details
A/1/1463
" be" in [context]
Zoom to ↗ View details
A/1/1648
" is" in [context]
Zoom to ↗ View details
A/1/1787
" are" in [context]
Zoom to ↗ View details
A/1/2307
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2341
" been" in scientific manuscripts (shown, demonstrated, proposed)
Zoom to ↗ View details
A/1/2539
" was"/" were" in biomedical studies
Zoom to ↗ View details
A/1/2542
" was"/" were" in medical contexts?
Zoom to ↗ View details
A/1/2844
" be" in [context]
Zoom to ↗ View details
A/1/3339
" is" in [context]
Zoom to ↗ View details
A/1/1223
"'m"
Zoom to ↗ View details
A/1/138
',"' or ",'" concluding what somebody said
Zoom to ↗ View details
A/1/2561
"."" in [context]
Zoom to ↗ View details
A/1/2591
[Ultralow density cluster]
Zoom to ↗ View details
A/0/334
The feature fires for the period (.) character followed by quotation marks (").
Zoom to ↗ View details
Cluster #115
A/0/47
The feature detects the use of quotation marks as it activates whenever it encounters paired single or double quotation marks.
Zoom to ↗ View details
A/1/1203
"."" in [context]
Zoom to ↗ View details
A/1/2346
"."" in [context]
Zoom to ↗ View details
A/1/2816
Closing punctuation+quote esp in series of quotes with no unquoted text
Zoom to ↗ View details
A/1/3421
"."" in [context]
Zoom to ↗ View details
A/1/2001
This feature fires when sentences end in question marks or exclamation points.
Zoom to ↗ View details
A/1/311
".”" (period + smart close quote)
Zoom to ↗ View details
Cluster #114
A/0/200
The feature seems to fire the most when a sentence ends with a period.
Zoom to ↗ View details
A/1/170
"." in sports discussion
Zoom to ↗ View details
A/1/1497
"." in [context]
Zoom to ↗ View details
A/1/2230
".'"
Zoom to ↗ View details
A/1/3932
"." in [context]
Zoom to ↗ View details
A/1/3968
"." in [context]
Zoom to ↗ View details
Cluster #0
A/0/25
Zoom to ↗ View details
A/0/89
Zoom to ↗ View details
A/0/120
Zoom to ↗ View details
A/0/124
Zoom to ↗ View details
A/0/208
Zoom to ↗ View details
A/0/299
Zoom to ↗ View details
A/0/309
Zoom to ↗ View details
A/0/336
Zoom to ↗ View details
A/0/467
Zoom to ↗ View details
A/1/31
[Ultralow density cluster]
Zoom to ↗ View details
A/1/37
[Ultralow density cluster]
Zoom to ↗ View details
A/1/57
[Ultralow density cluster]
Zoom to ↗ View details
A/1/79
[Ultralow density cluster]
Zoom to ↗ View details
A/1/82
[Ultralow density cluster]
Zoom to ↗ View details
A/1/95
[Ultralow density cluster]
Zoom to ↗ View details
A/1/122
[Ultralow density cluster]
Zoom to ↗ View details
A/1/131
[Ultralow density cluster]
Zoom to ↗ View details
A/1/135
[Ultralow density cluster]
Zoom to ↗ View details
A/1/168
[Ultralow density cluster]
Zoom to ↗ View details
A/1/188
[Ultralow density cluster]
Zoom to ↗ View details
A/1/190
[Ultralow density cluster]
Zoom to ↗ View details
A/1/191
[Ultralow density cluster]
Zoom to ↗ View details
A/1/214
[Ultralow density cluster]
Zoom to ↗ View details
A/1/217
[Ultralow density cluster]
Zoom to ↗ View details
A/1/223
[Ultralow density cluster]
Zoom to ↗ View details
A/1/227
[Ultralow density cluster]
Zoom to ↗ View details
A/1/238
[Ultralow density cluster]
Zoom to ↗ View details
A/1/240
[Ultralow density cluster]
Zoom to ↗ View details
A/1/242
[Ultralow density cluster]
Zoom to ↗ View details
A/1/250
[Ultralow density cluster]
Zoom to ↗ View details
A/1/260
[Ultralow density cluster]
Zoom to ↗ View details
A/1/266
[Ultralow density cluster]
Zoom to ↗ View details
A/1/286
[Ultralow density cluster]
Zoom to ↗ View details
A/1/288
[Ultralow density cluster]
Zoom to ↗ View details
A/1/298
[Ultralow density cluster]
Zoom to ↗ View details
A/1/299
[Ultralow density cluster]
Zoom to ↗ View details
A/1/309
[Ultralow density cluster]
Zoom to ↗ View details
A/1/312
[Ultralow density cluster]
Zoom to ↗ View details
A/1/318
[Ultralow density cluster]
Zoom to ↗ View details
A/1/325
[Ultralow density cluster]
Zoom to ↗ View details
A/1/354
[Ultralow density cluster]
Zoom to ↗ View details
A/1/361
[Ultralow density cluster]
Zoom to ↗ View details
A/1/368
[Ultralow density cluster]
Zoom to ↗ View details
A/1/383
[Ultralow density cluster]
Zoom to ↗ View details
A/1/387
[Ultralow density cluster]
Zoom to ↗ View details
A/1/401
[Ultralow density cluster]
Zoom to ↗ View details
A/1/414
[Ultralow density cluster]
Zoom to ↗ View details
A/1/428
[Ultralow density cluster]
Zoom to ↗ View details
A/1/433
[Ultralow density cluster]
Zoom to ↗ View details
A/1/440
[Ultralow density cluster]
Zoom to ↗ View details
A/1/446
[Ultralow density cluster]
Zoom to ↗ View details
A/1/448
[Ultralow density cluster]
Zoom to ↗ View details
A/1/459
[Ultralow density cluster]
Zoom to ↗ View details
A/1/468
[Ultralow density cluster]
Zoom to ↗ View details
A/1/490
[Ultralow density cluster]
Zoom to ↗ View details
A/1/504
[Ultralow density cluster]
Zoom to ↗ View details
A/1/518
[Ultralow density cluster]
Zoom to ↗ View details
A/1/562
[Ultralow density cluster]
Zoom to ↗ View details
A/1/568
[Ultralow density cluster]
Zoom to ↗ View details
A/1/574
[Ultralow density cluster]
Zoom to ↗ View details
A/1/581
[Ultralow density cluster]
Zoom to ↗ View details
A/1/602
[Ultralow density cluster]
Zoom to ↗ View details
A/1/637
[Ultralow density cluster]
Zoom to ↗ View details
A/1/659
[Ultralow density cluster]
Zoom to ↗ View details
A/1/670
[Ultralow density cluster]
Zoom to ↗ View details
A/1/695
[Ultralow density cluster]
Zoom to ↗ View details
A/1/710
[Ultralow density cluster]
Zoom to ↗ View details
A/1/719
[Ultralow density cluster]
Zoom to ↗ View details
A/1/742
[Ultralow density cluster]
Zoom to ↗ View details
A/1/779
[Ultralow density cluster]
Zoom to ↗ View details
A/1/793
[Ultralow density cluster]
Zoom to ↗ View details
A/1/801
[Ultralow density cluster]
Zoom to ↗ View details
A/1/886
[Ultralow density cluster]
Zoom to ↗ View details
A/1/892
[Ultralow density cluster]
Zoom to ↗ View details
A/1/926
[Ultralow density cluster]
Zoom to ↗ View details
A/1/950
[Ultralow density cluster]
Zoom to ↗ View details
A/1/966
[Ultralow density cluster]
Zoom to ↗ View details
A/1/977
[Ultralow density cluster]
Zoom to ↗ View details
A/1/987
[Ultralow density cluster]
Zoom to ↗ View details
A/1/992
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1006
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1011
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1018
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1048
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1058
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1071
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1072
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1097
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1109
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1118
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1121
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1151
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1158
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1161
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1170
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1217
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1221
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1228
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1247
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1252
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1257
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1265
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1267
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1289
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1290
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1294
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1324
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1331
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1336
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1341
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1356
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1370
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1393
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1402
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1407
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1421
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1427
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1468
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1486
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1492
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1504
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1516
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1523
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1525
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1528
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1530
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1540
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1569
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1574
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1585
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1590
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1594
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1617
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1646
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1684
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1702
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1705
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1710
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1711
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1715
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1734
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1744
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1748
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1756
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1765
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1767
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1769
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1792
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1816
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1817
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1819
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1831
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1839
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1854
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1861
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1869
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1873
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1884
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1888
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1897
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1899
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1946
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1963
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1966
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1971
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1976
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1989
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1994
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2004
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2010
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2029
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2043
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2050
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2080
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2086
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2093
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2097
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2110
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2120
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2128
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2138
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2218
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2220
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2223
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2238
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2246
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2247
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2257
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2297
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2305
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2333
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2355
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2365
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2388
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2392
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2408
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2437
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2441
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2447
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2470
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2475
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2480
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2490
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2511
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2521
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2537
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2541
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2544
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2551
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2564
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2575
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2581
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2583
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2619
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2643
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2648
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2682
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2685
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2692
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2714
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2716
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2717
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2726
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2733
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2749
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2771
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2775
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2833
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2858
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2869
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2870
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2891
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2895
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2896
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2912
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2923
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2924
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2935
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2936
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2938
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2961
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2962
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2966
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2980
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2987
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2991
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3025
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3051
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3069
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3079
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3086
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3092
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3101
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3103
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3113
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3117
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3127
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3148
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3153
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3168
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3170
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3171
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3172
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3187
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3202
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3221
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3239
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3251
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3257
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3266
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3268
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3273
This feature fires for words related to light, specifically luminescence.
Zoom to ↗ View details
A/1/3277
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3285
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3296
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3300
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3302
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3312
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3330
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3360
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3372
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3377
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3390
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3391
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3395
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3398
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3409
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3431
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3436
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3473
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3478
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3480
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3497
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3531
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3538
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3541
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3544
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3566
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3571
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3575
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3586
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3597
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3633
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3640
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3645
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3654
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3661
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3662
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3686
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3695
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3707
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3714
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3719
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3722
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3724
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3727
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3731
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3735
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3744
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3764
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3770
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3773
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3775
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3781
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3786
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3795
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3824
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3834
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3862
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3867
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3896
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3930
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3935
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3938
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3943
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3946
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3951
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3959
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3971
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3996
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4002
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4009
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4015
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4017
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4021
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4029
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4039
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4057
[Ultralow density cluster]
Zoom to ↗ View details
Cluster #104
A/0/15
This feature fires when the current token is followed by a period.
Zoom to ↗ View details
A/1/248
"." in [context]
Zoom to ↗ View details
A/1/264
"." ending sentences in historical/chronological contexts
Zoom to ↗ View details
A/1/415
Period (in media review?)
Zoom to ↗ View details
A/1/537
"." ending a sentence in a news report about politics
Zoom to ↗ View details
A/1/614
"." in [context]
Zoom to ↗ View details
A/1/913
"." in [context]
Zoom to ↗ View details
A/1/1514
Periods and other sentence-concluding punctuation, esp. in video games?
Zoom to ↗ View details
A/1/1722
"." in [context]
Zoom to ↗ View details
A/1/1755
"." in [context]
Zoom to ↗ View details
A/1/1809
"." in [context]
Zoom to ↗ View details
A/1/2013
Periods, esp. shortly after EOT
Zoom to ↗ View details
A/1/2294
"." in [context]
Zoom to ↗ View details
A/1/2411
"." ending sentences in business/finance news contexts?
Zoom to ↗ View details
A/1/2495
")." as end of English language parenthetical
Zoom to ↗ View details
A/1/3294
"." in [context]
Zoom to ↗ View details
A/1/3515
"." in [context]
Zoom to ↗ View details
A/1/3636
"." in [context]
Zoom to ↗ View details
A/1/3958
"." in [context]
Zoom to ↗ View details
Cluster #103
A/0/135
The feature fires when it sees punctuation marks and symbols like periods, parentheses, brackets, dollar signs and mathematical symbols at the end of tokens.
Zoom to ↗ View details
A/0/498
This feature fires when it sees punctuation at the end of a sentence or phrase, particularly periods "."
Zoom to ↗ View details
A/1/513
"." ending a sentence about medical care or formal medical reports
Zoom to ↗ View details
A/1/531
"." ending a sentence in a scientific paper, mostly materials science
Zoom to ↗ View details
A/1/532
"." ending a sentence in a journal paper or formal publication about disease or immunology or similar
Zoom to ↗ View details
A/1/831
Periods at the end of sentences in papers on radio astronomy
Zoom to ↗ View details
A/1/1357
"]\\]."
Zoom to ↗ View details
A/1/2368
End parenthetical + new sentence in academic param
Zoom to ↗ View details
A/1/3924
"." in [context]
Zoom to ↗ View details
A/1/4024
This feature fires when a sentence ends with a period.
Zoom to ↗ View details
A/1/796
" ." used in repeated " . . ." style contexts
Zoom to ↗ View details
A/1/3810
"." in [context]
Zoom to ↗ View details
Cluster #76
A/0/424
This feature fires when it sees a sentence ending in a question mark.
Zoom to ↗ View details
A/1/587
"!"
Zoom to ↗ View details
A/1/2770
"?" in news?
Zoom to ↗ View details
A/0/285
This feature attends to sentences that end with a period followed by whitespace.
Zoom to ↗ View details
A/1/621
"?" in dialogue?
Zoom to ↗ View details
A/1/3668
"." in [context]
Zoom to ↗ View details
Cluster #67
A/1/979
" ;)"
Zoom to ↗ View details
A/1/2062
" :)"
Zoom to ↗ View details
A/1/2855
"." in [context]
Zoom to ↗ View details
Cluster #72
A/0/46
This 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.
Zoom to ↗ View details
A/1/277
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1282
Ubuntu/linux conversations (irc chat?)
Zoom to ↗ View details
A/1/1902
" the" in installations, operating systems?
Zoom to ↗ View details
A/1/4079
Ubuntu/Linux/GNOME terms, esp. lowercase in chat
Zoom to ↗ View details
A/1/2641
(IRC?) usernames being mentioned/replied to chat logs, after > and before :
Zoom to ↗ View details
A/1/2035
Punctuation in paths or package names near "ubuntu"/"linux"/etc.
Zoom to ↗ View details
Cluster #71
A/0/151
This feature fires when it sees <> and : style chat messages, IRC style messages, and other messaging styles like that.
Zoom to ↗ View details
A/1/149
":" in ">"-based chat logs?
Zoom to ↗ View details
A/1/2616
">" in [context]
Zoom to ↗ View details
Cluster #82
A/0/474
This feature fires for instances of a single space in the token sequence.
Zoom to ↗ View details
A/1/1117
The first space in a double-space between sentences, in a legal context
Zoom to ↗ View details
A/1/3053
Newline+spaces near Old/poetic English?
Zoom to ↗ View details
A/1/3720
" " in [context]
Zoom to ↗ View details
A/1/3899
Line breaks followed by many spaces in legal contexts
Zoom to ↗ View details
A/1/4023
" " in [context]
Zoom to ↗ View details
A/1/50
"..." and some other long English punctuation
Zoom to ↗ View details
Cluster #101
A/0/437
This feature fires when it sees line breaks (like em-dashes, hyphens, vertical bars) used as separators, especially between sentences or clauses.
Zoom to ↗ View details
A/1/185
Closing parenthesis on years
Zoom to ↗ View details
A/1/598
"\u201d"
Zoom to ↗ View details
A/1/2548
""" in [context]
Zoom to ↗ View details
A/1/3559
"space em-dash" (em dashes typeset open) in English prose
Zoom to ↗ View details
A/1/3592
")" in [context]
Zoom to ↗ View details
A/1/2759
The feature attends to punctuation around proper nouns.
Zoom to ↗ View details
Cluster #66
A/0/367
This feature fires when it sees structured references (e.g., titles, sections, footnotes, page numbers) that are commonly seen in academic writing and publications.
Zoom to ↗ View details
A/1/960
Punctuation near ""-surrounded numbers/variables
Zoom to ↗ View details
A/1/1916
Continue chains of commas (and maybe related punctuation)?
Zoom to ↗ View details
Cluster #62
A/0/11
This feature seems to fire on numerical decimal points, especially those with 2-3 digits before and after the decimal point.
Zoom to ↗ View details
A/1/116
Legal text "." and ")." ?
Zoom to ↗ View details
A/1/498
[Ultralow density cluster]
Zoom to ↗ View details
A/1/569
Decimal point in a number
Zoom to ↗ View details
A/1/1114
"." as a decimal separator
Zoom to ↗ View details
A/1/1236
" ." in [context]
Zoom to ↗ View details
A/1/1286
"." in [context]
Zoom to ↗ View details
A/1/1821
"+"
Zoom to ↗ View details
A/1/2106
"." in [context]
Zoom to ↗ View details
A/1/2684
"." in [context]
Zoom to ↗ View details
A/1/2768
"." among long strings of digits, often ones with multiple "."s
Zoom to ↗ View details
A/1/2889
"." between numbers near lots of spaces
Zoom to ↗ View details
A/1/3082
"." in [context]
Zoom to ↗ View details
A/1/3609
"." interspersed between more than two numbers (version numbers, IP addresses; but not decimal numbers)
Zoom to ↗ View details
A/1/2130
" ?"
Zoom to ↗ View details
Cluster #100
A/0/507
The feature fires on the colon (:) symbol.
Zoom to ↗ View details
A/1/694
":" in [context]
Zoom to ↗ View details
A/1/959
" -" in [context]
Zoom to ↗ View details
A/1/1004
: separator in times
Zoom to ↗ View details
A/1/1914
":" in [context]
Zoom to ↗ View details
A/1/2229
":" in [context]
Zoom to ↗ View details
A/1/2347
":" in [context]
Zoom to ↗ View details
A/1/2713
: to indicate someone is speaking in a script
Zoom to ↗ View details
A/1/4091
":" in citations of sources
Zoom to ↗ View details
Cluster #97
A/1/408
":" after "\n\nA" (in Q: and A: pairs?)
Zoom to ↗ View details
A/1/2144
":" in [context]
Zoom to ↗ View details
A/1/2186
":" in [context]
Zoom to ↗ View details
A/1/2908
" !"
Zoom to ↗ View details
A/0/402
The 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.
Zoom to ↗ View details
A/1/96
":" esp. after "abstract" (in some paper metadata?)
Zoom to ↗ View details
A/1/2846
":" in [context]
Zoom to ↗ View details
A/1/3751
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3669
Punctuation or non-content links in forum posts?
Zoom to ↗ View details
A/1/2817
Time specification
Zoom to ↗ View details
A/1/1013
" "/" " in [context]
Zoom to ↗ View details
A/1/1537
Long spans of whitespace in poorly-formatted documents, often near decimal points
Zoom to ↗ View details
Cluster #86
A/0/426
This feature attends to formatted tabular data such as titles, tables, lists, tables of contents, etc.
Zoom to ↗ View details
A/1/1299
Spaces near vertical bars (Markdown/ASCII tables?)
Zoom to ↗ View details
A/1/3826
Runs of periods
Zoom to ↗ View details
A/1/4090
" |" in Markdown tables?
Zoom to ↗ View details
A/1/2885
Emoji ?
Zoom to ↗ View details
Cluster #85
A/1/91
ASCII/Unicode box drawings
Zoom to ↗ View details
A/1/972
Long runs of spaces (in technical writing / math?)
Zoom to ↗ View details
A/1/1717
Math
Zoom to ↗ View details
A/1/2139
"-->" indicating the span of time between two listed timestamps (and occasionally captures hexadecimal base pairs separated by spaces?)
Zoom to ↗ View details
A/1/682
Pipe-separated text (esp. words)
Zoom to ↗ View details
A/1/1852
LaTeX "bf" and other macros after ""
Zoom to ↗ View details
Cluster #45
A/0/95
The feature attends to hexadecimal numbers.
Zoom to ↗ View details
A/1/818
Comma-separated numbers
Zoom to ↗ View details
A/1/884
Digits after "0x" in "0x" hex literals
Zoom to ↗ View details
A/1/1399
" "/" "/" "/" "/" "/" " in [context]
Zoom to ↗ View details
A/1/2213
"," separating hexadecimal literals like "0xf1"
Zoom to ↗ View details
A/1/2430
" 0" beginning "0x" hexadecimal literals
Zoom to ↗ View details
A/1/3203
Hex numbers after "0x", usually in code defining ALLCAPS constants
Zoom to ↗ View details
A/1/2605
This feature fires on words that contain a dash "-", particularly scientific and biochemistry related terms.
Zoom to ↗ View details
A/1/3942
"-" in [context]
Zoom to ↗ View details
A/1/1272
"-" in [context]
Zoom to ↗ View details
Cluster #41
A/0/114
The feature fires when it encounters a hyphen followed by lowercase letters like in chemical compound names and hyphenated phrases.
Zoom to ↗ View details
A/1/99
Hyphen in Title-Case compounds
Zoom to ↗ View details
A/1/145
Hyphen after ALLCAPS abbreviation (SCPs, etc...)
Zoom to ↗ View details
A/1/2113
"-" connecting last names
Zoom to ↗ View details
A/1/2327
"-" in [context]
Zoom to ↗ View details
A/1/2569
"/" in [context]
Zoom to ↗ View details
Cluster #30
A/0/112
This feature fires on hyphens between numbers, especially when denoting age.
Zoom to ↗ View details
A/1/12
Hyphen, esp. after "Code § " and a number, denoting a legal code section
Zoom to ↗ View details
A/1/906
"/" for number/number in normal language text, often used for things like dates or simple fractions, but not in a math formula
Zoom to ↗ View details
A/1/1326
"-" in [context]
Zoom to ↗ View details
A/1/1678
"\u2013"
Zoom to ↗ View details
A/1/2115
"-" in identifiers/numbers/phone-numbers
Zoom to ↗ View details
A/1/2967
Seperator in dates
Zoom to ↗ View details
A/1/3712
^/^- in scientific notation
Zoom to ↗ View details
Cluster #40
A/0/243
This feature seems to be attending to hyphens followed by another word or number.
Zoom to ↗ View details
A/0/286
This feature fires on hyphenated words often indicating prefixes and suffixes.
Zoom to ↗ View details
A/1/246
"-" in hyphenated constructs esp. after "anti"
Zoom to ↗ View details
A/1/333
"-" after common compound-word-forming preposition: "in"/"on"/"over"/...
Zoom to ↗ View details
A/1/489
"-" in [context]
Zoom to ↗ View details
A/1/556
"-" in [context]
Zoom to ↗ View details
A/1/902
"-" in [context]
Zoom to ↗ View details
A/1/986
"-" in [context]
Zoom to ↗ View details
A/1/1435
"-" in [context]
Zoom to ↗ View details
A/1/1524
"-" in [context]
Zoom to ↗ View details
A/1/1887
"-" in [context]
Zoom to ↗ View details
A/1/3007
"-" after "well"
Zoom to ↗ View details
A/1/3110
"-" in [context]
Zoom to ↗ View details
A/1/3245
"-" in [context]
Zoom to ↗ View details
A/1/3271
"-" in [context]
Zoom to ↗ View details
A/1/3423
"-" in [context]
Zoom to ↗ View details
A/1/3474
"-" in "non-"
Zoom to ↗ View details
A/1/3643
"-" in [context]
Zoom to ↗ View details
A/1/3931
"-" in [context]
Zoom to ↗ View details
A/1/1043
"/" separating language concepts like "and/or" or "noun/othernoun"
Zoom to ↗ View details
A/1/2438
"-" in [context]
Zoom to ↗ View details
A/0/276
This feature attends to punctuation like hyphens, dashes, slashes, parentheses, quotes, apostrophes, etc.
Zoom to ↗ View details
A/1/1115
Hyphen used to join words
Zoom to ↗ View details
A/0/385
This feature seems to fire after underscore characters when they are being used to emphasize or highlight the words/phrases before and after them.
Zoom to ↗ View details
A/1/3679
"" typically concluding italics for a title, e.g. in Markdown
Zoom to ↗ View details
Cluster #29
A/0/157
The feature fires when there are quotation marks or other types of quotes around words or phrases.
Zoom to ↗ View details
A/1/19
Various European languages: Opening quotes?
Zoom to ↗ View details
A/1/639
" "" quoting some legal text
Zoom to ↗ View details
A/1/805
" \u201c" in [context]
Zoom to ↗ View details
A/1/927
" "" for opening an airquote or small-phrased-quote
Zoom to ↗ View details
A/1/1129
Opening quotes, mostly non-ascii and stronger on double-quotes
Zoom to ↗ View details
A/1/2819
"\u2014", emdash
Zoom to ↗ View details
A/1/3649
Opening quotes, esp. " '" and esp. after a comma
Zoom to ↗ View details
A/1/4084
" _"
Zoom to ↗ View details
Cluster #28
A/0/401
The feature fires when the token directly after a period is a quotation mark.
Zoom to ↗ View details
A/1/1763
"\u201c"
Zoom to ↗ View details
A/1/2701
" "" for someone speaking to follow a sentence of their behavior or previous speaking, as in a story
Zoom to ↗ View details
A/1/3123
" "" to as part of rapid-fire back-and-forth conversation
Zoom to ↗ View details
A/1/3728
" ""as part of longer back-and-forth conversation
Zoom to ↗ View details
A/0/53
This feature fires strongly on quotation marks, especially double quotation marks “ “.
Zoom to ↗ View details
A/1/194
'"' after two newlines
Zoom to ↗ View details
Cluster #132
A/0/4
This feature fires on line breaks and paragraphs especially in legal or technical documents.
Zoom to ↗ View details
A/1/2235
Code?: Double newlines in legal text or code?
Zoom to ↗ View details
A/1/3350
Quadruple newlines in legal text
Zoom to ↗ View details
A/1/3767
Triple newlines, often after EOT
Zoom to ↗ View details
A/1/3463
Newlines in recipies
Zoom to ↗ View details
Cluster #163
A/0/354
This feature attends to line breaks (new lines) in text, particularly double line breaks.
Zoom to ↗ View details
A/1/1693
Newline before numbers in legal text?
Zoom to ↗ View details
A/1/2853
Newline in legal text
Zoom to ↗ View details
A/1/3462
Newline in legal judgment?
Zoom to ↗ View details
Cluster #130
A/1/804
Court filings
Zoom to ↗ View details
A/1/1198
Newline+Form Feed in legal text?
Zoom to ↗ View details
A/1/1582
"\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]
Zoom to ↗ View details
Cluster #133
A/0/429
This feature fires for line breaks such as carriage returns and newlines.
Zoom to ↗ View details
A/1/184
Code: "\r\n\r" (carriage-return, newline, carriage-return)
Zoom to ↗ View details
A/1/257
Carriage returns, usu. before newlines, in legal contexts
Zoom to ↗ View details
A/1/625
Code: Carriage returns, usu. before newlines, esp. in C or similar languages
Zoom to ↗ View details
A/1/2443
Carriage returns, usu. followed by newlines, close to EOT and/or in HTML
Zoom to ↗ View details
A/1/2934
Carriage return
Zoom to ↗ View details
A/1/3947
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4068
Code: Carriage returns, usu. before newlines, in CSS
Zoom to ↗ View details
Cluster #167
A/0/100
This feature fires when there is a newline character "\n".
Zoom to ↗ View details
A/0/192
The feature fires when a new paragraph starts.
Zoom to ↗ View details
A/0/372
This feature fires for line breaks / new lines.
Zoom to ↗ View details
A/1/171
Newlines after proper nouns?
Zoom to ↗ View details
A/1/1598
Newline mid-sentence in art/fiction/nature?
Zoom to ↗ View details
A/1/1619
Newline shortly after EOT
Zoom to ↗ View details
A/1/2092
Newline in discussion of accounts/authentication
Zoom to ↗ View details
A/1/2259
Newline in economics?
Zoom to ↗ View details
A/1/2354
Newline after/between quotes/dialogue
Zoom to ↗ View details
A/1/2555
Newline in English prose (fiction? religion/philosophy?)
Zoom to ↗ View details
A/1/2615
Newlines in website updates, payment/subscription management
Zoom to ↗ View details
A/1/2727
Newlines in patents? (mechanical engineering?)
Zoom to ↗ View details
A/1/2734
Newline mid-sentence esp. in religious text
Zoom to ↗ View details
A/1/2760
Newlines before level-3 Markdown/ATX headers
Zoom to ↗ View details
A/1/2822
Newlines after punctuation that's not just a period
Zoom to ↗ View details
A/1/3141
Newline in news reports?
Zoom to ↗ View details
A/1/3155
Various European languages: Newline?
Zoom to ↗ View details
A/1/3236
Newline after sentence-ending period
Zoom to ↗ View details
A/1/3499
Newline in bibliography list?
Zoom to ↗ View details
A/1/3779
Newline after "was"/"were" mid-sentence
Zoom to ↗ View details
Cluster #181
A/0/106
This feature fires when there's a new line that follows an end punctuation mark (., !, ?, ", etc.).
Zoom to ↗ View details
A/1/148
Newline after another newline, predicting new sentence/item
Zoom to ↗ View details
A/1/2454
Newlines after "References"
Zoom to ↗ View details
A/1/948
Newline in chat logs? (IRC?)
Zoom to ↗ View details
Cluster #183
A/0/475
This feature fires on new line characters ("\n").
Zoom to ↗ View details
A/1/400
Newlines near numbers or number-heavy IDs
Zoom to ↗ View details
A/1/545
Newline in biology contexts and/or near repeated "-"/"=" as horizontal dividers
Zoom to ↗ View details
A/1/652
Newline after some curly-brace-based header and before "-"-based horizontal rules
Zoom to ↗ View details
A/1/824
Newlines after math expressions in LaTeX
Zoom to ↗ View details
A/1/863
" \n" after ######
Zoom to ↗ View details
A/1/2182
Newline in biology near section markers?
Zoom to ↗ View details
A/1/2738
Newline separating Wikipedia categories
Zoom to ↗ View details
Cluster #184
A/0/375
This feature fires for text containing numbers ending in question marks on new lines.
Zoom to ↗ View details
A/1/74
Newlines in wikitext tables?
Zoom to ↗ View details
A/1/591
Newline near "="-based horizontal rules
Zoom to ↗ View details
A/1/1261
"\n\n" after "add your own caption"
Zoom to ↗ View details
A/1/3222
Newline in LaTeX esp. after beginning a math environment
Zoom to ↗ View details
A/1/2813
BibTeX
Zoom to ↗ View details
A/1/651
Synthetic math: Newline before/after one-line numbers in synthetic math problems
Zoom to ↗ View details
A/1/3352
Short sequences of asterisks?
Zoom to ↗ View details
A/1/845
[Ultralow density cluster]
Zoom to ↗ View details
A/0/132
The feature attends to Wikipedia articles, firing most for category headings.
Zoom to ↗ View details
A/1/2861
Wikipedia categories
Zoom to ↗ View details
A/1/919
Quotes/newlines in .po/.pot files (https://www.drupal.org/community/contributor-guide/reference-information/localize-drupal-org/working-with-offline/po-and-pot-files)
Zoom to ↗ View details
A/1/2124
Category + :
Zoom to ↗ View details
A/1/1494
[Ultralow density cluster]
Zoom to ↗ View details
A/1/275
Code: "-" esp. in "UTF-8" and other Unicode encoding names
Zoom to ↗ View details
Cluster #172
A/0/364
This feature fires on blocks of legal or license text.
Zoom to ↗ View details
A/1/72
"," in several memorized license strings
Zoom to ↗ View details
A/1/643
" FITNESS" in code licenses (GPL?)
Zoom to ↗ View details
A/1/743
WHETHER in a contract
Zoom to ↗ View details
A/1/803
"Competing interests" phrases
Zoom to ↗ View details
A/1/832
" under" in software licenses
Zoom to ↗ View details
A/1/1437
Copyright dates (and sometimes copyright organizations)
Zoom to ↗ View details
A/1/1647
All rights reserved. (and some other source code licensing texts)
Zoom to ↗ View details
A/1/1834
" *" (and sometimes "//") for use in copyright license multiline c/c++ comments
Zoom to ↗ View details
A/1/2105
" the" in the GPL license?
Zoom to ↗ View details
A/1/2135
Lists of 'Category' in movie metadata?
Zoom to ↗ View details
A/1/2711
" COURT"
Zoom to ↗ View details
A/1/2720
BSD License
Zoom to ↗ View details
A/1/2985
" picked"
Zoom to ↗ View details
A/1/3005
" respect"
Zoom to ↗ View details
A/1/3037
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3219
" License" in the GPL
Zoom to ↗ View details
A/1/3349
Funding acknowledgements
Zoom to ↗ View details
A/1/3691
Activates on "Copyright", predicts the copyright symbol and dates
Zoom to ↗ View details
A/1/3845
"Click here for additional data file" and supplemental data more generally
Zoom to ↗ View details
A/1/3029
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1662
Code: XML namespace references
Zoom to ↗ View details
A/1/3263
Code: Error variables/handling in Go?
Zoom to ↗ View details
A/1/1800
DOIs
Zoom to ↗ View details
A/1/2189
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2514
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2840
Newline mid-sentence or near colons?
Zoom to ↗ View details
Cluster #54
A/0/360
The feature fires on parentheses, especially at the end of lines and sentences.
Zoom to ↗ View details
A/1/853
'). in jQuery invocation
Zoom to ↗ View details
Cluster #33
A/0/326
This feature fires on code snippets, specifically identifier names, keywords, braces, semicolons, and other syntactical elements.
Zoom to ↗ View details
A/1/1418
Tabs near uppercase letters (stock tickers? code?)
Zoom to ↗ View details
A/1/3830
"\t" in [context]
Zoom to ↗ View details
A/1/3872
"\t" in [context]
Zoom to ↗ View details
A/1/3988
Newline after "-"-based horizontal rules
Zoom to ↗ View details
A/1/940
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3192
"://"
Zoom to ↗ View details
A/0/178
This feature fires on Twitter usernames in angled brackets (< and >).
Zoom to ↗ View details
Cluster #204
A/0/414
The feature fires when it sees groups of characters in URLs that often represent file paths, folders, file types, domains and subdomain names.
Zoom to ↗ View details
A/1/1570
"@", usually in email addresses
Zoom to ↗ View details
A/1/1979
Code: '="' in HTML attributes, particularly <a href>
Zoom to ↗ View details
A/1/2683
"/" and some other separators in URLs
Zoom to ↗ View details
A/1/3338
Code: hyphens in hyphenated code identifiers
Zoom to ↗ View details
A/1/4043
Code: "/" in import-paths
Zoom to ↗ View details
Cluster #175
A/1/152
Code: ":" in CSS specifically
Zoom to ↗ View details
A/1/1208
Code: CSS alignment (predicting that a vertical/horizontal alignment follows)
Zoom to ↗ View details
A/1/1430
Code: hyphen in CSS
Zoom to ↗ View details
A/1/1935
Code: "-" in font spec in CSS
Zoom to ↗ View details
A/1/2545
Code: =" when setting width/height in html/css
Zoom to ↗ View details
A/1/3577
"="" in [context]
Zoom to ↗ View details
A/1/3509
Subdomains or domains in URLs (after "https://" etc.)
Zoom to ↗ View details
Cluster #216
A/0/20
This feature attends to opening and closing HTML tags.
Zoom to ↗ View details
A/1/121
Code: HTML angle brackets opening/closing lists
Zoom to ↗ View details
A/1/290
Code: HTML tag syntax that could/does enclose table-related tags
Zoom to ↗ View details
A/1/968
Code: HTML/XML beginnings of close tags "</", sometimes with preceding punctuation
Zoom to ↗ View details
A/1/1030
Starts of HTML tags: "<" or "</", sometimes with preceding punctuation
Zoom to ↗ View details
A/1/1645
HTML tag beginnings "<", sometimes with preceding punctuation, predicting "br"
Zoom to ↗ View details
A/1/3556
"</", sometimes with preceding punctuation, beginning HTML closing tags
Zoom to ↗ View details
A/1/4005
Code: "><" in XML?
Zoom to ↗ View details
A/1/4081
Code: HTML Angle Brackets before UPPERCASE
Zoom to ↗ View details
A/1/279
"<" in chat logs
Zoom to ↗ View details
Cluster #174
A/0/284
This feature fires on HTML, CSS, and other code-heavy syntax like indentation, brackets, semicolons, and hash symbols.
Zoom to ↗ View details
A/1/953
Code: CSS preprocessor?
Zoom to ↗ View details
A/1/1449
Code: Newline +/- spaces in CSS
Zoom to ↗ View details
A/1/2253
Code: image file path manipulation
Zoom to ↗ View details
A/1/2635
"" for snake_case within androidmanifest.xml files
Zoom to ↗ View details
A/1/3003
Code: CSS padding
Zoom to ↗ View details
A/1/3218
Code: Punctuation in HTML properties
Zoom to ↗ View details
Cluster #215
A/0/265
This 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.
Zoom to ↗ View details
A/1/607
Code: "</" in XML (with <string>, <key> tags)
Zoom to ↗ View details
A/1/1720
Code: Token ends with HTML open tag character
Zoom to ↗ View details
A/1/2540
Opening quotes in technical writing
Zoom to ↗ View details
A/1/3056
"<" in [context]
Zoom to ↗ View details
A/1/3955
"<" in [context]
Zoom to ↗ View details
A/1/2423
Code: Punctuation near packages and versions, esp. in Node
Zoom to ↗ View details
A/1/2416
Code: opening delimiters and/or quotes of any kind
Zoom to ↗ View details
A/1/1764
"=" in [context]
Zoom to ↗ View details
Cluster #231
A/0/133
The feature seems to fire on code in various programming languages, particularly around string literals and function or variable names.
Zoom to ↗ View details
A/0/177
This feature seems to fire most strongly for open/close parentheses.
Zoom to ↗ View details
A/0/409
This 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.
Zoom to ↗ View details
A/1/39
Code: opening quotes in data?
Zoom to ↗ View details
A/1/918
Code: punctuation near price calculations/processing?
Zoom to ↗ View details
A/1/1344
Code: Opening parenthesis in C-like for/if/while statements
Zoom to ↗ View details
A/1/1616
Code: "" used in snake_case to join multiple words together, usually in code contexts
Zoom to ↗ View details
A/1/2179
Code: usually some opening delimiter + some quote, often handling basic website user info (PHP/Perl?)
Zoom to ↗ View details
A/1/2224
"[" in [context]
Zoom to ↗ View details
A/1/2262
Code: " $" as variable sigil
Zoom to ↗ View details
A/1/2274
"="" in [context]
Zoom to ↗ View details
A/1/3093
Code: "(" calling a function/constructor
Zoom to ↗ View details
A/1/3864
" @" in [context]
Zoom to ↗ View details
A/1/3210
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3019
Code: Punctuation near file handling
Zoom to ↗ View details
A/0/168
This feature seems to fire on code snippets and paths, looking for things like class names, function names, file extensions, #includes, and library imports.
Zoom to ↗ View details
A/1/2087
Code: punctuation in authentication-handling (PHP?)
Zoom to ↗ View details
A/1/1335
[Ultralow density cluster]
Zoom to ↗ View details
A/1/925
[Ultralow density cluster]
Zoom to ↗ View details
A/1/339
"" within citation-ish things, often followed by a year-of-publication
Zoom to ↗ View details
A/1/2310
Code: punctuation near multiprocessing and scheduling
Zoom to ↗ View details
A/1/2708
"::" in [context]
Zoom to ↗ View details
A/1/1069
Code: time manipulation
Zoom to ↗ View details
A/1/172
Code: Network programming?
Zoom to ↗ View details
Cluster #232
A/0/10
This 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.
Zoom to ↗ View details
A/0/462
This feature detects hyphens between words that indicate compound terms.
Zoom to ↗ View details
A/1/1546
Code: :: token in C++ namespaces
Zoom to ↗ View details
A/1/1846
Code: C for low-level (circuit boards etc.) programming?
Zoom to ↗ View details
A/1/2231
"->"
Zoom to ↗ View details
A/1/2748
Code: Underscores in snake_case identifiers
Zoom to ↗ View details
A/1/3747
Code: Open parentheses, sometimes with more punctuation, in C-like languages
Zoom to ↗ View details
A/1/1669
"" for snake_case within mathematical variable names
Zoom to ↗ View details
A/1/1007
Code: security key references
Zoom to ↗ View details
A/0/199
This feature appears to fire on tokens related to computer networking, such as IP addresses, TCP, TLS, SSL, HTTPS, ports, sockets, servers, certs, URLs, etc.
Zoom to ↗ View details
A/0/216
Zoom to ↗ View details
A/1/1450
Code: mostly punctuation in maybe function metaprogramming?
Zoom to ↗ View details
A/1/28
Code: making/manipulating HTTP requests and responses
Zoom to ↗ View details
A/1/1709
Code: " __" in special identifiers
Zoom to ↗ View details
A/1/3270
Words related to logging in/authentication
Zoom to ↗ View details
Cluster #141
A/0/358
This 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.
Zoom to ↗ View details
A/1/126
Code: Underscores in UPPER_SNAKE_CASE identifiers
Zoom to ↗ View details
A/1/209
[Ultralow density cluster]
Zoom to ↗ View details
A/1/477
Code: "" in UPPER_SNAKE_CASE identifiers (registers, low-level C?)
Zoom to ↗ View details
A/1/896
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1201
Code: CMake
Zoom to ↗ View details
A/1/2328
" "" for opening (and sometimes "." for closing) an all-caps quoted sentence, that guides further all-caps
Zoom to ↗ View details
A/1/3347
Code: _ in C++ macro, especially after BOOST
Zoom to ↗ View details
A/1/2444
Code: JavaScript or web frameworks (React etc.?)
Zoom to ↗ View details
A/1/575
Code: CSS
Zoom to ↗ View details
A/1/2009
Code: Command lines / shell scripts
Zoom to ↗ View details
A/1/2473
Code: C++? Gtk/UI programming?
Zoom to ↗ View details
A/1/3100
[Ultralow density cluster]
Zoom to ↗ View details
A/1/644
Code: C++ include statements
Zoom to ↗ View details
A/1/371
English descriptions of character classes
Zoom to ↗ View details
Cluster #118
A/0/207
This feature seems to fire on code-like tokens such as variable names, function names, keywords, preprocessor directives, nested blocks, etc.
Zoom to ↗ View details
A/1/1102
Code: ALL_CAPS with underscores
Zoom to ↗ View details
A/1/1325
ALLCAPS Syscall names, predicting runs of spaces?
Zoom to ↗ View details
A/1/1455
Code: Uppercase letters in SNAKE_CASE identifiers
Zoom to ↗ View details
A/1/2133
Code: Uppercase letters in SNAKE_CASE identifiers (in C preprocessor?)
Zoom to ↗ View details
A/1/3485
Capitalized letters and punctuation near/predicting "P"
Zoom to ↗ View details
A/1/3589
This feature seems to attend to proper names or acronyms starting with the letter M.
Zoom to ↗ View details
A/1/3651
TitleCase word prefixes predicting -ex, esp. in code contexts, esp. "Mut"
Zoom to ↗ View details
Cluster #194
A/0/257
This feature fires when it sees web addresses and URLs.
Zoom to ↗ View details
A/0/262
This 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 (~). Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-262) A/1/44 Usernames in a chat transcript of some sort? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-44) A/1/105 Words in URLs (excluding the "/" delimiters), focusing more on characters than numbers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-105) A/1/360 "com" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-360) A/1/540 Code: Various kinds of markup code (emails, html, cssy-stuff, misc data formats, etc) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-540) A/1/1458 TLDs in URLS (com in github.com, org in wikipedia.org) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1458) A/1/2089 Parts of domain names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2089) A/1/2620 Path components after and predicting "/" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2620) Cluster #198 A/1/289 "https" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-289) A/1/2370 Code: Linux command line Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2370) A/1/3011 " lib" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3011) A/1/3129 "http" in urls Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3129) Cluster #148 A/0/384 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-384) A/1/769 OCR Errors Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-769) A/1/2774 Text or word fragments, maybe medical Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2774) A/1/3010 " extends" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3010) A/1/3797 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3797) A/1/3085 Nouns in computer programming, which maybe hold code/state Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3085) A/1/2121 Object-oriented programming (English prose) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2121) A/1/938 Capitalized New Word Prediction Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-938) A/0/150 This feature fires on variable, function and class names in a variety of programming languages. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-150) A/1/1263 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1263) A/1/3367 Code: identifier fragments esp. non-English, or sometimes typos Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3367) Cluster #202 A/0/256 The feature fires when it sees code comments, denoted by /* ... */ or // in programming languages such as C, C++, Java, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-256) A/1/34 Code: Comments: Nouns in systems/database engineering? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-34) A/1/577 Code: documentation in markdown? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-577) A/1/1600 Code: Capitalized words at beginnings of code comments Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1600) A/1/1672 Code-related English prose, e.g. API documentation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1672) A/1/3382 Documentation around memory/CPU concepts (stack, pointer, structure, clock, size, etc) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3382) Cluster #201 A/0/362 This feature attends to code snippets, specifically Python code. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-362) A/1/751 Generic part of variable name Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-751) A/1/3848 Code: Names being imported in Python Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3848) A/1/3884 Code: Python imports Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3884) A/1/4089 Components of period-separated attribute/package paths Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4089) A/1/2147 Code: C style comments about low-level hardware Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2147) A/1/3444 TitleCase identifier components, esp. final ones, esp. with some abstract code meaning like "Class" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3444) A/1/692 Code: predicting byte/buffer? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-692) A/1/724 "include" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-724) A/1/567 "import" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-567) Cluster #65 A/0/386 This feature fires for numbers enclosed with symbols such as parentheses or brackets. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-386) A/1/1642 Caesar-shift encoded words Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1642) A/1/2950 ASCII control characters Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2950) A/1/2389 "x" in "0x" hex literals with many leading 0s Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2389) Cluster #107 A/0/119 This feature attends to hexadecimal strings. It appears to fire most strongly on strings of length 6-8. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-119) A/1/97 Long strings of digits Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-97) A/1/198 Lowercase hexadecimal (digit token) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-198) A/1/1444 Hexadecimal, esp. lowercase letters, predicting digits Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1444) A/1/2025 Digits esp. after the decimal point in decimals Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2025) A/1/2048 List of Numbers (space-separated) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2048) A/1/3629 "x" in hexadecimal literals Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3629) A/1/3817 Hexadecimal (uppercase) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3817) Cluster #64 A/0/45 This feature fires when it sees base64 encoded text. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-45) A/1/1493 Synthetic math: repetitive letter strings being picked "without replacement from" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1493) A/1/1544 base64 of ASCII Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1544) A/1/2357 base64 (esp. uppercase letters) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2357) A/1/2364 base64 digits Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2364) A/1/1961 Letters in short fragments of uppercase hexadecimal or some alphanumeric IDs Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1961) Cluster #106 A/0/283 The feature attends to numbers containing hyphens commonly found in phone numbers, dates, times, and item numbers. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-283) A/1/327 Hour parts of timestamps, esp. with leading space and 0 Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-327) A/1/1881 Numbers in various contexts e.g. error IDs Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1881) A/1/2215 Grant and project IDs, mostly digits with few capital letters Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2215) A/1/3807 2-digit time components separated by ":"s Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3807) A/1/4027 Numbers in hyphen-separated sequences Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4027) A/1/3941 Parts of phone numbers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3941) A/1/1036 Digits in URLs Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1036) A/1/2859 Dot-separated version numbers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2859) A/1/1573 File path components esp. related to installable software, packages, fonts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1573) A/1/605 Space+number tokens in various technical contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-605) A/1/3920 Code: Punctuation esp. at end of first line of class declarations Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3920) A/1/2945 Code: Objective C method declarations, with ":(" indicating a type will come next as part of a declaration Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2945) A/1/3020 Python-adjacent text/punctuation, esp. "#," preceding " python-format" in .po/.pot files Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3020) A/1/2998 Usernames in irc-style <username> format Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2998) A/0/153 This feature attends to HTML tags. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-153) Cluster #150 A/0/0 This feature attends to HTML and XML tags. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-0) A/1/173 Code: HTML attribute names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-173) A/1/259 Code: Numbers/hexadecimals inside quoted strings for UX-related config/markup (XML?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-259) A/1/978 Code: HTML class name, particularly abstract layout-related like "container" or "child" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-978) A/1/985 Android XML layout file properties Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-985) A/1/1082 Code: HTML/XML tag names in closing tags Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1082) A/1/1713 Code: Java/C++ types inside a template/generic instantiation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1713) A/1/3457 Code: HTML/XML opening tag names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3457) A/1/4031 " class" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4031) A/1/4086 HTML tag names (generally after "<") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4086) A/1/1879 Text in HTML or XML tags (predicting close tags) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1879) A/1/3034 Code: Parts of strings in compact JSON Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3034) A/1/4069 HTML/CSS tutorials (prose) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4069) Cluster #151 A/0/266 This feature attends to HTML programming code including HTML tags and attributes. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-266) A/1/244 Mid-word quoted text, predicting closing quotes Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-244) A/1/1124 Double-quotes string as first argument to a function or method call Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1124) A/1/2175 Code: Single-quoted strings (often in an array, set, or function call) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2175) A/1/3226 Words in string literals in JSON or similar in package metadata ("integrity", "metadata", "outputs") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3226) A/1/1603 Code: jQuery methods Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1603) A/1/2095 Short text in double quotes, esp. in technical contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2095) A/1/815 " click" / instructions for interacting with UIs Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-815) Cluster #192 A/0/209 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-209) A/1/1618 Letter strings that predict "ID"/"id" (no space) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1618) A/1/2060 Verbs and adjectives in discussions of tabular databases / spreadsheets Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2060) A/1/2164 Code: Testing-related Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2164) A/1/3518 Code: Visual Basic for Excel Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3518) A/1/3550 Code: SQL-related C/C++/Java code, constants, and comments Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3550) A/1/4010 Code: SQL Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4010) A/1/3825 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3825) A/1/2038 Code: Error conditions esp. "Failed" in error message string literals Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2038) A/1/1242 " var" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1242) A/1/3012 Code: .Net Manifest XML? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3012) A/1/1269 Code: related to IDs, users, accounts, and similar Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1269) A/1/2779 Code: Java CamelCase identifier constituents? ("Manager" "Application" "Bean" etc.) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2779) A/1/3072 Code: Field names in JavaScript Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3072) Cluster #240 A/0/39 The feature attends to code-like structures such as identifiers, keywords and symbols and longer code-fragments in various programming languages. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-39) A/0/211 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-211) A/1/236 Code: "get" in getter method names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-236) A/1/668 Code: CamelCase identifier components esp after "get" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-668) A/1/1026 Code: File Manipulation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1026) A/1/1140 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1140) A/1/1318 Code: the start of a line of indented code that is operating on a object/class (vs control flow) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1318) A/1/1749 Code: Java class/interface names in a class definition Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1749) A/1/1766 "nbsp" and some other HTML ampersand escape names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1766) A/1/1844 Code: Date/time-related in Java/Objective-C languages Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1844) A/1/1876 Code: Android UX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1876) A/1/2015 Code: CamelCase versions of UI/UX concepts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2015) A/1/2107 Code: Objective C function/method calls on iOS Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2107) A/1/2532 Code: relating to processes and tasks and async execution Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2532) A/1/2607 Code: words in CamelCase, particularly prepositions: "To", "For", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2607) A/1/2886 Code: C# components of CamelCase identifiers (metaprogramming?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2886) A/1/4012 Spanish and Portuguese: Pseudocode or coding related words? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4012) A/1/2892 Code: abstract-ish identifiers in object names, method calls, variable declarations Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2892) A/1/3228 Code: $-sigil-prefixed variable names (PHP/Perl?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3228) A/1/232 Code: some serialization or dump of networking info? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-232) A/1/1606 "Exception" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1606) A/1/1227 Code typically near string concatenation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1227) A/1/2417 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2417) Cluster #243 A/0/388 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-388) A/1/20 Code: println and other method calls, esp in Java Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-20) A/1/3165 Code: PHP loop iteratees? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3165) A/1/3684 Code: arguments in comma-separated argument lists Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3684) A/1/3644 Code: " String" in C++ Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3644) A/1/2342 Code: C++ " private" and other keywords Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2342) A/1/1149 Code: Variable names after "var", likely JavaScript Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1149) Cluster #173 A/0/511 This feature attends to code and programming language syntax such as variable declarations and function calls. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-511) A/1/186 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-186) A/1/454 Code: punctuation near geometry / graphics calculations Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-454) A/1/1762 Code: identifier components used in geometry / graphics Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1762) A/1/2306 " E" or other identifier prefixes or punctuation predicting "GL" in OpenGL contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2306) A/1/4080 This feature attends to text dealing with lung diseases, specifically chronic obstructive pulmonary disease (COPD). Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4080) A/1/1995 Code: Components of CamelCase identifiers (in C/C++?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1995) A/1/1476 Code: Python (Machine Learning) (esp. punctuation) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1476) A/1/2193 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2193) A/0/349 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-349) A/1/2574 Code: Tree traversal (mix of Java, Go, Javascript, maybe others) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2574) Cluster #225 A/1/430 " char" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-430) A/1/837 " void" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-837) A/1/1845 " unsigned"/"unsigned" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1845) A/1/3286 " int" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3286) A/1/2101 Code: identifier fragments, often after "_", related to data or dimension/shape Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2101) A/0/460 This feature attends to programming language syntax, particularly C/C++. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-460) A/1/386 Code: C++ template type variables, esp. " T" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-386) A/1/573 Code: " std" and other C++ namespaces (anticipating "::") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-573) Cluster #207 A/0/140 The feature appears to fire on identifier names in programming languages. It attends to variable, function, and class names in code. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-140) A/0/411 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-411) A/1/29 Code: Beginnings of components of snake_case identifiers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-29) A/1/218 Code: short variable names? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-218) A/1/447 " i" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-447) A/1/1051 Code: space+letter tokens, often single-letter variable names) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1051) A/1/2993 Code: C in hardware drivers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2993) A/1/144 Code: C++ type-level programming Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-144) A/1/94 Code: Python Definition Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-94) A/1/3992 Code: " null" or sometimes " false" esp. after ":" or "return" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3992) A/1/1760 Fields accessed with member-of-pointer (arrow operator "->") in C++ Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1760) A/1/2468 " len" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2468) A/1/899 " x" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-899) A/1/3048 Code: argument name after "(" or "(&" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3048) Cluster #92 A/0/305 This feature seems to fire on numbers with decimal points, especially when they are close to 0.5. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-305) A/1/705 Code: Single digits, usually not numeric literals (e.g. in identifiers or versions) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-705) A/1/718 Single-digit tokens esp. after "=", often some software flag or argument Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-718) A/1/3989 " 0" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3989) Cluster #242 A/0/196 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-196) A/1/409 Code: Identifiers after "(", likely arguments, in C#? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-409) A/1/1729 Code: variable in if () condition Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1729) A/1/3033 Code: words referring to measurements or quantities: "length"/"width"/"height"/"count" etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3033) A/1/3128 Code: End of parenthetical Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3128) A/1/1442 Code: Small numeric constants in parentheses (predicting close parentheses) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1442) A/1/1379 " As" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1379) Cluster #265 A/0/282 This 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". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-282) A/1/16 " than" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-16) A/1/3849 " than" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3849) A/0/351 This feature attends to the word "as" used as a conjunction. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-351) Cluster #264 A/0/104 This feature fires when it sees the word "as" being used as a preposition or conjunction. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-104) A/1/555 " as" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-555) A/1/631 " as" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-631) A/1/1978 " as" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1978) A/1/3252 " as" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3252) A/1/3666 " as" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3666) A/1/963 Adverb in "as ... as" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-963) A/1/1083 "As" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1083) A/1/1390 " since" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1390) A/1/3293 " until" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3293) A/1/1686 " except" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1686) A/1/3674 This feature fires when transition words like "furthermore", "moreover", "thus", etc are used. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3674) A/1/3035 " while" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3035) Cluster #253 A/1/1111 Express uncertainty with maybe/probably/perhaps Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1111) A/1/1784 " then" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1784) A/1/2054 " Well" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2054) Cluster #274 A/0/502 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-502) A/1/622 " but" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-622) A/1/1560 Sentence-starting word like "Moreover" and "Therefore" (predicting ",") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1560) A/1/3525 " however" mid-sentence (after commas or semicolons) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3525) Cluster #275 A/0/29 This feature fires when the word "if" is present. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-29) A/1/592 " if" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-592) A/1/2278 "if" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2278) A/1/2406 "If" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2406) A/1/3059 "though"/"although"/"while"/forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3059) A/1/3852 " Whether"/" whether" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3852) A/1/3918 " when" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3918) Cluster #263 A/0/116 This feature fires in response to transition words like "however", "but", "although", "though", "even though", "yet", and "until". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-116) A/0/191 The feature fires for the word "And" as a discourse or stylistic marker to conjoin phrases or clauses. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-191) A/1/355 " Or"/"Or" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-355) A/1/1021 " but" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1021) A/1/2088 " Nor" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2088) A/1/2972 Sentence-starting conjunctions " And"/" But" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2972) A/1/3304 " So"/"So" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3304) A/1/2531 " because" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2531) A/1/2699 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2699) Cluster #269 A/0/141 This feature attends to phrases that contain the word "that". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-141) A/1/254 " that" Relative Pronoun, news Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-254) A/1/450 " that" Relative Pronoun, Scientific Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-450) A/1/604 " that" Relative Pronoun, Narrative Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-604) A/1/1177 " that" conjuction scientific Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1177) A/1/1250 " that" Indicative/Conjunction Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1250) A/1/1551 " that" Demonstrative Pronoun Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1551) A/1/1863 " which" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1863) A/1/2645 " that" legal Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2645) A/1/4082 " which" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4082) A/0/127 This feature appears to activate when surrounding text contains relative pronouns like "which", "who", and "whom". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-127) Cluster #258 A/0/222 This feature attends to the interrogative pronoun "what". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-222) A/1/372 " how" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-372) A/1/752 "Why" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-752) A/1/2102 " what" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2102) A/1/2603 " What"/"What" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2603) A/1/3555 "How" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3555) A/1/1362 " That"/"That" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1362) A/1/1484 " Oh"/"Oh" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1484) A/1/501 " who" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-501) A/1/263 " Which"/"Which" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-263) A/1/3288 Affirmative utterances: "Yeah", "Yes", "Okay" etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3288) Cluster #177 A/0/186 This feature is firing for interrogative sentences starting with “what”. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-186) A/1/130 " can"/" could"/" would" after "How" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-130) A/1/156 Word after "What" where "what" is the start of a phrase Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-156) A/1/1580 " you" as a question (often a yes/no question) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1580) A/1/2122 Concepts/words in questions, predicting the end of the question "?", optionally with more punctuation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2122) A/1/3626 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3626) Cluster #170 A/0/122 The feature attends to statements or phrases using the words "there is," "there are," or "here" to indicate existence, presence, location or emphasis. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-122) A/0/394 This feature fires when the token is "it's" or "It's". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-394) A/1/141 " it"/" It" in scientific contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-141) A/1/600 " it" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-600) A/1/618 Abstract nouns in "If"/"When" sentences, esp. computing or astrophysics related (e.g. "algorithm", "universe") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-618) A/1/794 " it" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-794) A/1/1610 " you" and other pronouns after "thought" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1610) A/1/1776 " it" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1776) A/1/1870 "It" in "It's", at the start of quoted speech Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1870) A/1/2168 " There"/"There" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2168) A/1/2546 " there" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2546) A/1/3152 " It"/"It" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3152) A/1/3460 " Here"/"Here" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3460) A/1/3366 Emphasis, specificity, or exclusivity Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3366) Cluster #189 A/0/24 This feature attends to first person plural pronouns. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-24) A/0/73 This feature fires when the word "you" or variations of it ("your", "you've", etc.) are present. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-73) A/0/171 The feature fires when it sees second person pronouns such as "you", "your", and "yours". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-171) A/0/190 This feature attends to sentences with the pronoun "they", especially in contexts where "they" refers to people. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-190) A/1/77 " authors" (and capitalization/spacing variants) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-77) A/1/486 " We"/" we" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-486) A/1/616 " You"/"You" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-616) A/1/655 " they" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-655) A/1/1218 " You"/"You" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1218) A/1/1346 " you" in conditional "if" sentences Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1346) A/1/1793 " They"/" they" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1793) A/1/2084 " we", in political punditry / media reviews? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2084) A/1/2270 " we", esp. after a comma, in legal contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2270) A/1/2493 "We" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2493) A/1/2568 Words after "we"/"We", esp. adverbs indicating time or chronological position ("firstly", "then", etc.) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2568) A/1/2827 "they"/"you"/"didn" predicting U+2019 Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2827) A/1/2847 " it"/" she"/" he" after "if" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2847) A/1/3014 " you" as subject of subclause Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3014) A/1/3310 " you" as direct object Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3310) A/1/3532 " We" in scientific contexts (royal "we found" etc.) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3532) Cluster #190 A/0/154 This feature attends to the pronoun "she" and associated female pronouns (her, herself). Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-154) A/0/482 The feature fires when it sees pronouns such as "he", "she", "who", etc. in the context of the sentence. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-482) A/1/650 " he"/" she" esp. in legal reports Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-650) A/1/745 " He"/"He" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-745) A/1/1342 " She"/" she"/"she" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1342) A/1/2069 "he" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2069) A/1/2549 This feature attends to instances of the pronoun "he", referring to a male person. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2549) Cluster #179 A/0/224 The feature detects first person singular pronouns and contractions such as "I", "I'm", "I've", "I'd", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-224) A/0/271 This feature attends to the usage of the word "I" and related pronouns like "I'm", "I've", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-271) A/1/52 " I" present tense Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-52) A/1/1725 " I" (or "I") at the beginning of a double-quoted statement Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1725) A/1/2241 " I", often in technical/troubleshooting contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2241) A/1/2624 "I" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2624) A/1/2640 " i", typically as a pronoun or in "iirc" in casual writing or chat Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2640) A/1/2715 " I" in a conditional(subjunctive?) sense: "when I", "If I", "because I", etc Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2715) A/1/2765 " I" at the beginning of a sentence Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2765) A/1/3487 " I", esp. in medical contexts as either the Roman numeral for 1 or start of an abbreviation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3487) A/1/3676 'I' usually after 'but' Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3676) A/1/4022 " I" in a proper noun (usually? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4022) A/1/4045 " I" past tense Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4045) A/1/874 " Which"/"Which" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-874) Cluster #223 A/0/444 This feature appears to fire for sentences with mathematical questions or formulas that reference specific terms in sequences, such as the nth term. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-444) A/1/201 Synthetic math: " replacement" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-201) A/1/1350 This feature attends to mathematical terms and concepts such as highest common factor, derivatives, prime factors, and integer rounding. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1350) A/1/1388 " nearest" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1388) A/1/2681 " prime" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2681) A/1/3682 " divided" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3682) A/1/3922 Synthetic math: " first"/" second"/" third" before " derivative" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3922) A/1/3167 Newlines before/after "True/False" in math problems Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3167) A/1/2834 Synthetic math: " is" in division problems Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2834) A/1/653 Numerical place ('thousands', 'tens', 'hundreds'), predicts digit, from math dataset Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-653) A/1/3397 " the" in "What is the..." etc. math problems Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3397) A/1/3280 "-" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3280) A/1/735 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-735) A/1/4007 Synthetic math: ":" in multisets that are being sampled "without replacement from" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4007) Cluster #224 A/0/96 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-96) A/1/485 " to" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-485) A/1/777 "of" in a Math context, for denominator/term/derivative "of" a number Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-777) A/1/1813 " What"/"What" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1813) A/1/2609 " is" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2609) A/1/2764 " without" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2764) A/1/3554 " prob" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3554) A/1/3652 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3652) Cluster #185 A/1/930 " (" for multiple choice answers in math; generic "(" in math in lower quantiles Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-930) A/1/1323 " ||" in wikitext tables involving dates Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1323) A/1/2214 Synthetic math: Spaces and closing parentheses before choices in math multiple-choice questions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2214) A/1/2729 "⁄" U+2044 in normalized UTF-8 as latin1 mojibake Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2729) A/1/3602 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3602) A/1/2192 " and" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2192) Cluster #203 A/0/505 This feature fires on arithmetic symbols such as the minus sign and equals sign. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-505) A/1/1986 " +" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1986) A/1/2216 "," in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2216) A/1/2308 " -" as minus sign in math Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2308) A/1/2349 " -" indicating a negative number in a math problem Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2349) A/1/2474 This feature seems to fire when mathematical operations and notation are present, such as determining values, calculating probabilities, solving equations, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2474) A/1/2500 " =" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2500) A/1/3991 Verbs, esp. imperative, in (often synthetic) math Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3991) A/1/3693 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3693) Cluster #191 A/0/195 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-195) A/1/782 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-782) A/1/3343 "?" after non-words? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3343) A/1/3969 "." in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3969) A/1/2856 " be" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2856) Cluster #186 A/0/105 This feature attends to numerical formatting such as numbers, mathematical symbols and fractions. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-105) A/1/1826 "](#" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1826) A/1/2789 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2789) A/1/3188 Burmese: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3188) Cluster #219 A/0/293 This feature seems to attend to mathematical equations and formulas containing variables, operations, and numbers. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-293) A/1/2277 Synthetic math: "Let" introducing a math equation after a period Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2277) A/1/2753 " Suppose" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2753) A/1/2207 ")." as an end of a math formula sentence Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2207) A/1/14 Synthetic math: Open parenthesis in "Let a(b)..." Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-14) A/1/3865 Synthetic math: numbers being exponentiated (after "**") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3865) A/1/1022 "sqrt" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1022) A/1/1568 "**" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1568) Cluster #220 A/0/149 This feature appears to fire for numbers and math symbols, particularly exponents, in the context of mathematical equations and expressions. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-149) A/0/233 This feature attends to the phrases like 'let x = ...' and other statements of equations, variables, and mathematical expressions. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-233) A/1/62 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-62) A/1/495 Synthetic math: Single-letter variables or functions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-495) A/1/812 Single-letter variables in synthetic math esp. polynomials Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-812) A/1/847 Synthetic math: numeric constants before "*" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-847) A/1/855 This feature attends to mathematical expressions involving variables. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-855) A/1/1471 Synthetic math: numbers, about to be multiplied, after "Solve"/"Suppose" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1471) A/1/2302 Synthetic math: digits between newlines (problem numbers?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2302) A/1/2572 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2572) A/1/2662 Synthetic math: space followed by digit usu. after "+" or "-" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2662) A/1/3032 Year in Month Day Year format dates Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3032) A/1/3091 Synthetic math: numbers after single-letter keys "x:" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3091) A/1/3387 Synthetic math: number after "Let x =" (where x is any single letter) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3387) A/1/3400 LaTeX math-mode-ending "$" esp. shortly after "Let"? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3400) A/1/3526 Single-letter variables in polynomials? (where exponentiation is **) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3526) A/1/3568 This feature attends to basic arithmetic operations and equations involving addition, subtraction, multiplication and equal signs. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3568) A/1/3573 ")" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3573) A/1/3584 Synthetic math: single-letter variables esp. after number and "*" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3584) A/1/3762 Synthetic math: Single Token "*" in problems Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3762) Cluster #206 A/0/68 This feature appears to fire on mathematical expressions involving fractions, exponents, parentheses, variables, operators, and numbers. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-68) A/0/324 This feature fires on mathematical expressions and equations involving multiplication, exponentiation and variable names. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-324) A/1/46 "*" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-46) A/1/70 Synthetic math: numbers esp. in sqrt() Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-70) A/1/528 "(-" as the beginning of parenthesized negative number in math Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-528) A/1/583 Synthetic math: Variable name inside the parens of a definition of a function in math Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-583) A/1/593 Synthetic math: "(" of a sqrt(number) call where number is numbers and not a variable Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-593) A/1/781 Synthetic math?: Numbers and variables in equations Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-781) A/1/1321 ")/(-" or "/(-" in mathematical expressions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1321) A/1/1689 ")/" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1689) A/1/2743 Newline-space in some serialization format? (YAML?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2743) A/1/2952 Synthetic math: exponentiation starting with "**" with optional "(" and "-" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2952) A/1/3095 Synthetic math: minus sign in mathematical expressions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3095) A/1/3276 ")*(-" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3276) A/1/3472 "/" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3472) A/1/3617 Synthetic math: numbers in nested fractions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3617) A/1/3743 Synthetic math: Expression after "Simplify" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3743) A/1/3870 "/" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3870) A/1/332 xAB hex escapes (without backslashes or other delimiters) near text Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-332) A/1/842 LaTeX punctuation after \sum or \prod (either subscript or "\" anticipating "limits") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-842) A/1/1564 Consecutive numbers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1564) Cluster #210 A/0/292 This feature fires for mathematical expressions involving integers. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-292) A/1/315 Synthetic math: Arithmetic, numbers with small digits, in unusual bases Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-315) A/1/876 Numbers, esp. negative, in comma-separated lists Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-876) A/1/1599 Synthetic math: Numbers that we will need to answer whether it is composite/prime or divides other numbers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1599) A/1/2019 Synthetic math: Numbers or fractions that are to be sorted Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2019) A/1/2489 Synthetic math: problem involving division Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2489) A/1/2867 Synthetic math: Numbers in math problems Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2867) A/1/3306 Mathematical sequences (numbers separated by ", ") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3306) A/1/3354 Synthetic math: Numbers in math problems involving rounding Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3354) A/1/1441 "}{-" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1441) A/1/3814 Code: fileID, maybe numeric values in "key:value" more generally Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3814) A/1/635 Double-pipe delimited data Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-635) A/1/1416 "**" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1416) A/1/810 Numbers near Japanese and/or " -->" arrows Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-810) A/1/3305 Bug ids (mostly in URLs but also outside of URLs in context too) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3305) A/1/3076 Numbers written in exponential notation (but not the "E", which is predicted) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3076) A/1/2687 Markdown footnotes Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2687) Cluster #134 A/0/301 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-301) A/1/693 "######" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-693) A/1/728 Lots of dashes separating sections in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-728) A/1/2180 "================================" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2180) A/1/2298 Legal discovery emails, esp. "@" in "ECT@ECT" (Enron emails) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2298) A/1/2761 "Introduction" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2761) A/1/3341 " A/1/3514 "References" / other headers surrounded by (double) newlines Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3514) Cluster #102 A/0/110 This feature appears to attend to numbers. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-110) A/1/10 Numbers after 3-ish newlines (legal contexts?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-10) A/1/2504 1–2 digit tokens, esp. after double newlines, sometimes after other punctuation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2504) Cluster #135 A/0/347 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-347) A/1/1513 EOT at start Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1513) A/1/2754 EOT in middle Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2754) A/1/1033 Long runs of hyphens" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1033) A/1/1807 US court case numbers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1807) Cluster #128 A/0/87 This 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.
Zoom to ↗ View details
A/0/323
This 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.
Zoom to ↗ View details
A/1/717
Q before a : at the start of a sequence
Zoom to ↗ View details
A/1/1726
Title-Case metadata-ish nouns like "Title", "Name", or "Location"
Zoom to ↗ View details
A/1/2740
"0" or some other numbers after <EOT>
Zoom to ↗ View details
A/1/4036
"External"
Zoom to ↗ View details
A/1/2163
US Court of Appeals
Zoom to ↗ View details
A/1/377
Legal docs: Citation numbers in square brackets
Zoom to ↗ View details
A/1/2721
Legal docs
Zoom to ↗ View details
A/1/449
Period in legal docs, especially conflict of interest?
Zoom to ↗ View details
Cluster #123
A/1/1927
".", sometimes with preceding punctuation, in legal contexts
Zoom to ↗ View details
A/1/2404
"." in [context]
Zoom to ↗ View details
A/1/3231
"." in "v." (legal cases)
Zoom to ↗ View details
Cluster #96
A/0/405
The feature fires when sentences end with punctuation (e.g. period, question mark, exclamation mark).
Zoom to ↗ View details
A/1/1653
"." in [context]
Zoom to ↗ View details
A/1/2311
Period in mathematical proofs
Zoom to ↗ View details
A/1/2584
"." in [context]
Zoom to ↗ View details
A/1/3425
"." in [context]
Zoom to ↗ View details
A/1/3036
"." in [context]
Zoom to ↗ View details
Cluster #122
A/0/250
This feature fires when it sees case law citations.
Zoom to ↗ View details
A/0/346
This feature fires on abbreviations that are followed by periods such as "S." and "e.g.".
Zoom to ↗ View details
A/1/125
"." after "Dr" and few other honorifics
Zoom to ↗ View details
A/1/741
"." in [context]
Zoom to ↗ View details
A/1/1061
"." in [context]
Zoom to ↗ View details
A/1/1206
"." in [context]
Zoom to ↗ View details
A/1/1432
"." in [context]
Zoom to ↗ View details
A/1/1548
"." in [context]
Zoom to ↗ View details
A/1/1697
"." in abbreviations for legal citations
Zoom to ↗ View details
A/1/1783
"." in legal/formal abbreviations?
Zoom to ↗ View details
A/1/2373
"." in [context]
Zoom to ↗ View details
A/1/2642
"." in [context]
Zoom to ↗ View details
A/1/3520
"." after single letters, likely abbreviations ("i.e.", "e.g.", "p.")
Zoom to ↗ View details
A/1/3950
"." in [context]
Zoom to ↗ View details
A/1/4035
"." in [context]
Zoom to ↗ View details
A/1/2794
"." after initials of people (in academic contexts / citations?)
Zoom to ↗ View details
Cluster #91
A/0/376
This feature attends to citations of journal publications in the format "<Last name>, <Initials.>, <Publication name>, <Volume number> (<Year>)."
Zoom to ↗ View details
A/1/1060
"." in [context]
Zoom to ↗ View details
A/1/1343
".," in [context]
Zoom to ↗ View details
A/1/1438
")." suffixing a year, in a bibliography/citation context
Zoom to ↗ View details
A/1/1608
"]{},"
Zoom to ↗ View details
A/1/1640
" ["
Zoom to ↗ View details
A/1/2079
"." in abbreviations of chemistry/physics journals?
Zoom to ↗ View details
A/1/3796
"." in [context]
Zoom to ↗ View details
Cluster #69
A/1/1045
" \u00a7"
Zoom to ↗ View details
A/1/1125
Numbers inside of [NUM]{} citations
Zoom to ↗ View details
A/1/2533
Page numbers in citations
Zoom to ↗ View details
A/1/2865
" \u00b6"
Zoom to ↗ View details
Cluster #78
A/1/2173
" See"
Zoom to ↗ View details
A/1/3138
Patent Numbers
Zoom to ↗ View details
A/1/4006
" also" of "See also"
Zoom to ↗ View details
Cluster #81
A/0/22
This feature attends to references to sections and subsections of legal documents, such as statutes, patent numbers, court cases, etc.
Zoom to ↗ View details
A/0/169
This feature seems to attend to text that shows dates and days of the week, often in news-like contexts such as headlines.
Zoom to ↗ View details
A/1/61
Legal: ends of case numbers
Zoom to ↗ View details
A/1/136
Legal numbers after " at"
Zoom to ↗ View details
A/1/183
" v" in "v." (legal cases)
Zoom to ↗ View details
A/1/789
Numbers in legal citations, usually after "," and predicting a court
Zoom to ↗ View details
A/1/1489
Section number (after "section" or "§", esp. after "Code")
Zoom to ↗ View details
A/1/1559
"d"
Zoom to ↗ View details
A/1/1659
Legal digit esp. after "F." before "d"
Zoom to ↗ View details
A/1/1761
Supreme Court case numbers
Zoom to ↗ View details
A/1/1806
" Id"
Zoom to ↗ View details
A/1/2132
" Cir" Legal Citation
Zoom to ↗ View details
A/1/2222
This feature fires on parenthetical citations with a year inside such as (2002).
Zoom to ↗ View details
A/1/2706
Months in dates, predicting days of months
Zoom to ↗ View details
A/1/3749
Space and digit after "X v. Y," court case citations
Zoom to ↗ View details
A/1/3755
" supra"
Zoom to ↗ View details
A/1/4059
Numerical day of a date (Month day, day Month, Month day'th, etc)
Zoom to ↗ View details
A/1/398
"." in [context]
Zoom to ↗ View details
A/1/787
"." in "Fig." (predicting a figure number)
Zoom to ↗ View details
A/1/1612
"." in [context]
Zoom to ↗ View details
A/1/1512
Supplementary data in papers
Zoom to ↗ View details
Cluster #94
A/0/378
This feature fires on the label "Fig." that usually precedes a figure number in academic papers.
Zoom to ↗ View details
A/1/278
" Fig"
Zoom to ↗ View details
A/1/1531
" Table"
Zoom to ↗ View details
A/1/2488
" FIG"
Zoom to ↗ View details
A/1/3723
" Eq" before an equation reference
Zoom to ↗ View details
Cluster #95
A/0/238
This feature attends to citations of figures in academic papers.
Zoom to ↗ View details
A/1/265
Short table names (" S" in "Table S1.")
Zoom to ↗ View details
A/1/282
""}"
Zoom to ↗ View details
A/1/362
"Fig"
Zoom to ↗ View details
A/1/3070
Figure numbers (after "Figure"/"Fig"/...)
Zoom to ↗ View details
A/1/370
"[", especially after "Fig. "
Zoom to ↗ View details
A/1/543
Letters in figure names that are parenthesized or bracketed, like "[Figure 4A]"
Zoom to ↗ View details
A/1/3675
Many hyphens
Zoom to ↗ View details
A/1/353
Chemical molecular data, showing atoms, bonds, and geometry, near many spaces and "---"
Zoom to ↗ View details
A/1/283
"/" as dimensional units "x per y"
Zoom to ↗ View details
Cluster #129
A/0/465
This feature attends to lines and horizontal lines in particular.
Zoom to ↗ View details
A/1/2051
Many hyphens
Zoom to ↗ View details
A/1/3868
Whitespace, often right after EOT
Zoom to ↗ View details
A/1/3880
Long runs of spaces in medical reports?
Zoom to ↗ View details
Cluster #26
A/0/160
This 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.
Zoom to ↗ View details
A/1/473
"[" in [context]
Zoom to ↗ View details
A/1/1229
" (**"
Zoom to ↗ View details
A/1/2949
" [" in [context]
Zoom to ↗ View details
A/1/3291
" (["
Zoom to ↗ View details
A/1/702
{ and ^{ in LaTeX
Zoom to ↗ View details
Cluster #25
A/0/229
This feature attends to parentheses () in the text.
Zoom to ↗ View details
A/1/457
" (" in [context]
Zoom to ↗ View details
A/1/536
Opening parenthesis after name or institution, where the parenthesis often contains a date
Zoom to ↗ View details
A/1/929
Open parentheses after more than one newline
Zoom to ↗ View details
A/1/1065
" (" in an informal narrative or commentary
Zoom to ↗ View details
A/1/1259
" (" Citation, Reference, or Note
Zoom to ↗ View details
A/1/1491
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2276
" (" medical/epidemiological context with a focus on statistical analysis
Zoom to ↗ View details
A/1/2943
" (" in a math/physics technical paper
Zoom to ↗ View details
A/1/3064
"(" in a legal context referencing a particular clause: ie "924(c)(1)"
Zoom to ↗ View details
A/1/3365
" (" in a legal/judicial context
Zoom to ↗ View details
A/1/3449
" (" in [context]
Zoom to ↗ View details
A/1/3689
Open parentheses with no leading space, often in math notation or chemical formulas
Zoom to ↗ View details
A/1/3776
" (" in a medical/biological context with a focus on biological terms/parenthetical-acronyms
Zoom to ↗ View details
A/1/3844
" (" in a software-development/technical document
Zoom to ↗ View details
Cluster #21
A/0/52
This feature attends to acronyms in parentheses.
Zoom to ↗ View details
A/0/249
The feature attends to parentheticals, which are phrases or clauses set off by parentheses or other punctuation marks.
Zoom to ↗ View details
A/1/157
Letter after opening parenthesis, particularly "(i.e.)" or "(e.g.)"
Zoom to ↗ View details
A/1/292
Parenthesized numbers of quantities/units
Zoom to ↗ View details
A/1/768
Short lowercase word after "(", esp. in English prose mid-sentence
Zoom to ↗ View details
A/1/1457
Single lower case letter in parentheses
Zoom to ↗ View details
A/1/2204
"see"
Zoom to ↗ View details
A/1/2248
Years in parentheses, esp. citations
Zoom to ↗ View details
A/1/2296
Words inside parenthetical expressions that could be the end of the parenthetical phrase, predicting punctuation including ")"
Zoom to ↗ View details
A/1/2553
Text in parentheses, esp. proper nouns
Zoom to ↗ View details
A/1/2673
Parenthesized abbreviations (abbreviating the preceding text?)
Zoom to ↗ View details
A/1/2842
This feature activates most strongly when it sees the presence of mathematical functions and equations, signified by characters like parentheses, exponents, division signs, etc.
Zoom to ↗ View details
A/1/3088
"0" in [context]
Zoom to ↗ View details
A/1/3096
"," in [context]
Zoom to ↗ View details
A/1/3466
Words inside parentheses, esp. right after open parentheses, in medical/experimental contexts
Zoom to ↗ View details
A/1/4074
Numbered citations/footnotes
Zoom to ↗ View details
A/1/3990
" \u00b1"
Zoom to ↗ View details
Cluster #138
A/0/313
This feature attends to numbers with decimal points representing measurements such as weight or concentration.
Zoom to ↗ View details
A/1/2140
" **" in [context]
Zoom to ↗ View details
A/1/3162
"--"
Zoom to ↗ View details
Cluster #140
A/0/472
This feature fires for numbers formatted with spacing around the decimal points.
Zoom to ↗ View details
A/1/59
" \u23ce\n "/" \u23ce\n "/"\u23ce\n "/"\u23ce\n " in [context]
Zoom to ↗ View details
A/1/166
Long runs of whitespace near numbers/statistics
Zoom to ↗ View details
A/1/404
" "/" "/" "/" "/" "/" "/" " in [context]
Zoom to ↗ View details
A/1/726
Runs of whitespace in medical contexts?
Zoom to ↗ View details
A/1/1178
" "/" "/" "/" "/" "/" " in [context]
Zoom to ↗ View details
A/1/1453
Long spaces next to escaped arithmetic operators, potentially in tables?
Zoom to ↗ View details
A/1/3028
" =" in [context]
Zoom to ↗ View details
A/1/3808
" ×" in numeric dimension pairs
Zoom to ↗ View details
A/1/2377
"---"
Zoom to ↗ View details
A/1/1382
" "/" "/" "/" "/" " in [context]
Zoom to ↗ View details
A/1/3504
Capital letters and/or punctuation predicting 3-digit numbers?
Zoom to ↗ View details
Cluster #137
A/0/88
This feature seems to fire on numbers, punctuation like parentheses, commas, periods, colons, and mathematical operators like +, -, =.
Zoom to ↗ View details
A/1/88
Closing parenthesis after number predicting run of space
Zoom to ↗ View details
A/1/3567
Punctuation or tokens preceding "fn" referring to footnotes in {} citation format
Zoom to ↗ View details
A/1/134
" vs" / "versus" etc.
Zoom to ↗ View details
A/1/133
Hebrew?: mojibake UTF-8 as latin1?
Zoom to ↗ View details
A/1/712
" " when used in sequence with additional "" and " * tokens, in comments
Zoom to ↗ View details
A/1/3040
" ", usually as markdown bold delimiters
Zoom to ↗ View details
Cluster #139
A/0/400
The feature attends to tables and formatting, specifically numbers, text alignment, and punctuation.
Zoom to ↗ View details
A/1/1424
Numbers and punctuation near numbers, often "%)"
Zoom to ↗ View details
A/1/2424
Human body or medical supply related, near and predicting long runs of spaces
Zoom to ↗ View details
A/1/2781
" \<"
Zoom to ↗ View details
A/1/3885
"" in [context]
Zoom to ↗ View details
A/0/477
This 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.
Zoom to ↗ View details
A/1/819
"," in [context]
Zoom to ↗ View details
Cluster #136
A/0/67
The feature fires when it sees numbers.
Zoom to ↗ View details
A/1/1510
Digits, usually 2 or 3, after the decimal point in such numbers separated by many spaces
Zoom to ↗ View details
A/1/3467
" 0" in [context]
Zoom to ↗ View details
A/1/3756
Predicting whitespace after legal citations
Zoom to ↗ View details
Cluster #89
A/0/247
The feature fires when it sees dates formatted as two-digit to four-digit numbers representing years between 1900-2020.
Zoom to ↗ View details
A/1/869
A year, in talking about history and what happened
Zoom to ↗ View details
A/1/1634
Years, 19xx or 20xx, often in citations of journals and the like
Zoom to ↗ View details
A/1/642
"" in [context]
Zoom to ↗ View details
A/1/834
ANOVA and related statistical tests
Zoom to ↗ View details
A/1/66
[Ultralow density cluster]
Zoom to ↗ View details
Cluster #131
A/0/94
The feature attends to numbers with decimals in scientific writing.
Zoom to ↗ View details
A/0/242
This feature fires when it sees statistical symbols and formatting, particularly p-values, standard error, and statistical significance indicators.
Zoom to ↗ View details
A/1/469
Numbers in percentages
Zoom to ↗ View details
A/1/709
Digits after the decimal point after "=" in parentheses
Zoom to ↗ View details
A/1/2793
Numbers after/near " ±"
Zoom to ↗ View details
A/1/3233
"%" in [context]
Zoom to ↗ View details
A/1/3887
" C" in [context]
Zoom to ↗ View details
Cluster #90
A/0/261
This feature attends to numbers between 20-50.
Zoom to ↗ View details
A/1/224
" million"/" billion" big numbers or things measured in such quantities
Zoom to ↗ View details
A/1/235
Various European languages: Number text
Zoom to ↗ View details
A/1/740
Page numbers esp. in page ranges
Zoom to ↗ View details
A/1/828
Numbers after Title-Case nouns suggesting sections of a larger work, like "Volume" or "Issue"
Zoom to ↗ View details
A/1/867
Single letters, esp. lowercase, immediately after numbers
Zoom to ↗ View details
A/1/954
Inline numbers in English text, esp. >2 digits predicting measurements/units
Zoom to ↗ View details
A/1/1126
This feature fires on numbers that contain a 1, 4 or 8 in the tenths or hundredths place.
Zoom to ↗ View details
A/1/1209
Digits in mixed digits/uppercase letter sequences (uppercase base36 maybe)
Zoom to ↗ View details
A/1/1862
Small single-digit numbers (as Arabic numerals)
Zoom to ↗ View details
A/1/1992
Numbers in English sentences
Zoom to ↗ View details
A/1/2026
~2-digit numbers (ages, page numbers)
Zoom to ↗ View details
A/1/2448
Numbers in Bible verse citations (separated by colons)
Zoom to ↗ View details
A/1/2634
This feature attends to numbers greater than or equal to 100.
Zoom to ↗ View details
A/1/2820
Numbers after " in"/" on" etc., predicting months
Zoom to ↗ View details
A/1/3703
Numeric dollar amounts, after "$"
Zoom to ↗ View details
A/1/3902
This feature fires when it sees numbers and letter combinations that appear to be vehicle, product or model numbers.
Zoom to ↗ View details
A/1/908
"" in [context]
Zoom to ↗ View details
A/1/226
"^" in Markdown-style superscripts in chemistry?
Zoom to ↗ View details
Cluster #110
A/0/152
This feature attends to numerical values and measurements.
Zoom to ↗ View details
A/1/1183
Nuclear Magnetic Resonance
Zoom to ↗ View details
A/1/1954
" 0" in [context]
Zoom to ↗ View details
A/1/2100
Space + 2 or 3 digit number, inline in English prose (mechanical engineering?)
Zoom to ↗ View details
A/1/2337
~2-digit numbers in medical studies
Zoom to ↗ View details
A/1/2394
Numbers of units after the decimal point
Zoom to ↗ View details
A/1/2812
" 1"
Zoom to ↗ View details
A/1/2857
Numbers before acronyms/units in scientific papers
Zoom to ↗ View details
A/1/3284
1 to 3 digit numbers, often scientific measurements/units
Zoom to ↗ View details
A/1/2210
"~" in chemistry Markdown subscripts
Zoom to ↗ View details
A/1/533
This feature fires when it sees exponents, superscripts, and other mathematical notation.
Zoom to ↗ View details
Cluster #111
A/0/51
The feature fires when it detects scientific notation and chemical formulae.
Zoom to ↗ View details
A/1/154
Badly formatted math operators
Zoom to ↗ View details
A/1/816
~2-digit numbers in scientific contexts, predicting various units
Zoom to ↗ View details
A/1/1465
Numbers in Markdown ("~") subscripts near single capital letters (chemistry?)
Zoom to ↗ View details
A/1/2167
Code: double-asterisk-emphasized numbers and letters
Zoom to ↗ View details
A/1/2362
" \u00b0"
Zoom to ↗ View details
A/1/1162
" (" Preceding Roman Numeral
Zoom to ↗ View details
A/1/461
Code: HTML & escape character before lt/gt/nbsp/quot/amp
Zoom to ↗ View details
A/1/2877
Punctuation in JSON-like data, regexes... predicting single digits
Zoom to ↗ View details
Cluster #39
A/0/466
This feature attends to the ampersand symbol (&).
Zoom to ↗ View details
A/1/393
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4013
LaTeX table (delimiters?)
Zoom to ↗ View details
Cluster #38
A/0/407
This feature fires on numeric values, including both integers and decimals.
Zoom to ↗ View details
A/1/388
Numbers in LaTeX tables
Zoom to ↗ View details
A/1/878
Whitespace esp. near/between LaTeX math mode expressions
Zoom to ↗ View details
A/1/1122
Base-ten numbers, especially decimal, negative, and in comma-separated lists without spaces
Zoom to ↗ View details
A/1/2351
Numbers and some English/abbreviations, separated by and predicting & (LaTeX?)
Zoom to ↗ View details
A/1/2393
Numbers in (X,Y) pairs? (coordinate-based plotting?)
Zoom to ↗ View details
A/1/1670
Punctuation tokens in math ending with "(" and near "x"
Zoom to ↗ View details
A/1/1991
Math notation, synthetic math operators, or LaTeX near the letter "w"
Zoom to ↗ View details
A/0/287
This feature attends to LaTeX code blocks.
Zoom to ↗ View details
A/1/488
Newline-spaces in LaTeX commands
Zoom to ↗ View details
Cluster #63
A/0/146
This feature attends to LaTeX math expressions containing fractions.
Zoom to ↗ View details
A/0/228
This feature seems to fire for dollar signs ($) used in mathematical formulas and equations.
Zoom to ↗ View details
A/0/303
This feature attends to mathematical expressions, especially those containing Greek letters, math symbols, and formulas.
Zoom to ↗ View details
A/0/340
This feature attends to subscripts that use curly braces such as C{D}.
Zoom to ↗ View details
A/0/470
This feature attends to subscripts of mathematical variables.
Zoom to ↗ View details
A/0/501
This feature attends to mathematical equations and formulas that contain exponents.
Zoom to ↗ View details
A/1/167
"^{"
Zoom to ↗ View details
A/1/237
LaTeX: "-" in math mode (subtraction) esp. superscripts or subscripts
Zoom to ↗ View details
A/1/330
LaTeX: " =" assigning a variable
Zoom to ↗ View details
A/1/345
"{" in LaTeX after mbox/hbox/text/...
Zoom to ↗ View details
A/1/364
Synthetic math: operators near "f"
Zoom to ↗ View details
A/1/444
"{" in [context]
Zoom to ↗ View details
A/1/813
"(" and other opening delimiters after LaTeX "\left"
Zoom to ↗ View details
A/1/931
Backslashes predicting "u" in Unicode character escapes
Zoom to ↗ View details
A/1/1089
Synthetic math: parentheses and arithmetic near "c"
Zoom to ↗ View details
A/1/1181
"{" in [context]
Zoom to ↗ View details
A/1/1233
LaTeX begin inline math
Zoom to ↗ View details
A/1/1256
Predicting Short Math Variable
Zoom to ↗ View details
A/1/1328
"{" in [context]
Zoom to ↗ View details
A/1/1754
Math punctuation near/predicting "n"
Zoom to ↗ View details
A/1/1768
LaTeX subscripts with preceding punctuation, predicting "i", "j", etc.
Zoom to ↗ View details
A/1/1923
" $" in [context]
Zoom to ↗ View details
A/1/1955
LaTeX: "" used for subscript
Zoom to ↗ View details
A/1/1982
" $" in [context]
Zoom to ↗ View details
A/1/2263
LaTeX: Underscores and similar punctuation, predicting digits
Zoom to ↗ View details
A/1/2316
"^" in [context]
Zoom to ↗ View details
A/1/2405
"," in [context]
Zoom to ↗ View details
A/1/3180
"}{" in [context]
Zoom to ↗ View details
A/1/3248
Synthetic math: "" in polynomials etc
Zoom to ↗ View details
A/1/3403
"{" in [context]
Zoom to ↗ View details
A/1/3529
Open parentheses in LaTeX?
Zoom to ↗ View details
A/1/3569
Parens, brackets, subscripts, and mathematical operators, predicting a variable "k"
Zoom to ↗ View details
A/1/3632
Mostly capital letter fragments predicting X, Y, Z
Zoom to ↗ View details
A/1/3889
"{" in [context]
Zoom to ↗ View details
A/1/3952
Operators and punctuation in math near/predicting "q"
Zoom to ↗ View details
A/1/3976
"=" in [context]
Zoom to ↗ View details
A/1/871
Letters near "+", often but not always scientific notation "1e+5"
Zoom to ↗ View details
A/1/1017
LaTeX esp. opening braces
Zoom to ↗ View details
Cluster #58
A/0/416
This feature fires on math equations written in LaTeX.
Zoom to ↗ View details
A/1/375
LaTeX: Curly brace in setting mathbb font
Zoom to ↗ View details
A/1/399
"}{" in LaTeX, esp. array environments
Zoom to ↗ View details
A/1/678
"{" in [context]
Zoom to ↗ View details
A/1/1664
"{" in [context]
Zoom to ↗ View details
A/1/1105
Punctuation in LaTeX
Zoom to ↗ View details
A/1/1275
LaTeX measurements
Zoom to ↗ View details
A/1/1135
Capital letters or punctuation near/predicting "G"
Zoom to ↗ View details
A/1/1947
"}" in [context]
Zoom to ↗ View details
A/1/107
LaTeX math environment names
Zoom to ↗ View details
Cluster #34
A/0/312
This 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.
Zoom to ↗ View details
A/0/316
This feature attends to LaTeX math notation.
Zoom to ↗ View details
A/1/55
"}" in [context]
Zoom to ↗ View details
A/1/273
"]{" in LaTeX
Zoom to ↗ View details
A/1/679
LaTeX: " \"/"\" in [context]
Zoom to ↗ View details
A/1/943
"}" in [context]
Zoom to ↗ View details
A/1/1076
"}" in [context]
Zoom to ↗ View details
A/1/1205
"}" in [context]
Zoom to ↗ View details
A/1/2381
"12"
Zoom to ↗ View details
A/1/2778
"}" in [context]
Zoom to ↗ View details
A/1/3490
"{" in [context]
Zoom to ↗ View details
A/1/3981
"was"
Zoom to ↗ View details
Cluster #47
A/0/329
The feature attends to LaTeX math environments, equations, and symbols.
Zoom to ↗ View details
A/0/418
This feature fires for LaTeX math mode environments, equations, theorems, tables, figures, algorithms, etc.
Zoom to ↗ View details
A/1/113
" ["
Zoom to ↗ View details
A/1/1132
LaTeX equation tags
Zoom to ↗ View details
A/1/1283
Citations and references after (\[, predicting \])
Zoom to ↗ View details
A/1/1461
" (\\["
Zoom to ↗ View details
A/1/2205
Filenames in maybe Markdown link targets?
Zoom to ↗ View details
A/1/3543
"{" esp. in LaTeX labels
Zoom to ↗ View details
A/1/3578
" \\["
Zoom to ↗ View details
A/1/3587
":" in [context]
Zoom to ↗ View details
A/1/3660
Markup references to tables/algorithms/etc., in brackets before colons
Zoom to ↗ View details
Cluster #27
A/0/21
This feature fires when it sees brackets [] around numbers, letters and variable names.
Zoom to ↗ View details
A/1/396
Code: Single-digit numbers in square-bracket array indexing
Zoom to ↗ View details
A/1/3505
English in [square brackets]
Zoom to ↗ View details
A/1/2094
"$-", sometimes with prior punctuation, joining hyphenated LaTeX-math-and-word forms
Zoom to ↗ View details
A/1/212
"." surrounding sub/sup for subscript/superscript formatting requests
Zoom to ↗ View details
A/1/359
Data Label prefixes: basically of the form: .[]{data-label="
Zoom to ↗ View details
A/1/999
LaTeX tokens after open square brackets (left ends of intervals, etc.)
Zoom to ↗ View details
A/1/3275
LaTeX formatting
Zoom to ↗ View details
A/1/3863
"amsmath"
Zoom to ↗ View details
A/1/851
Code: identifiers (variables, key or field names) in curly braces
Zoom to ↗ View details
Cluster #292
A/0/259
This feature fires when it sees mathematical symbols and equations such as dollar signs, subscripts, fractions, greek letters, parentheses, etc.
Zoom to ↗ View details
A/1/402
"}$$"
Zoom to ↗ View details
A/1/638
This feature fires when equations (written in LaTeX or MathML format) are present in the text.
Zoom to ↗ View details
A/1/1507
"\\]"
Zoom to ↗ View details
A/1/2266
Closing LaTeX math mode
Zoom to ↗ View details
A/1/3912
This feature attends to mathematical symbols and formulas like variables, operators, functions, fractions etc.
Zoom to ↗ View details
A/1/1475
"$" in [context]
Zoom to ↗ View details
Cluster #291
A/0/40
The feature attends to mathematical notation such as inline formulas, equations, theorems, proofs, symbols, Greek letters, subscripts, and other math syntax.
Zoom to ↗ View details
A/1/175
" Let" in/near math or LaTeX
Zoom to ↗ View details
A/1/879
" Then"/" then" in [context]
Zoom to ↗ View details
A/1/3557
Prepositions near LaTeX math
Zoom to ↗ View details
A/1/3873
" that" LaTeX Scientific
Zoom to ↗ View details
A/1/3646
Numeric tokens, sometimes after all-caps abbreviations/identifiers
Zoom to ↗ View details
A/1/1474
'follows' in logic, usually related to code
Zoom to ↗ View details
A/1/549
" such" in [context]
Zoom to ↗ View details
A/1/437
Mathematical quantifier, e.g. " some"/" any"/" every" after "for", near LaTeX and predicting " $"
Zoom to ↗ View details
A/1/1195
This feature attends to mathematical formulas.
Zoom to ↗ View details
A/1/1268
" Proposition"/" Theorem"/capitalized words for mathematical statements
Zoom to ↗ View details
A/1/936
Prepositions and relational verbs, esp. at beginnings of sentences in LaTeX/math-related contexts, predicting " Eq" or math mode starting
Zoom to ↗ View details
Cluster #307
A/0/281
This feature attends to mathematical terminology and notation related to abstract algebra, especially homomorphisms, isomorphisms, and topological spaces.
Zoom to ↗ View details
A/1/491
Topology, abstract algebra
Zoom to ↗ View details
A/1/1661
Advanced physics (mostly GR, but some QFT)
Zoom to ↗ View details
A/1/2595
Abstract adjectives in math: " convex", " regular", " arbitrary", etc.
Zoom to ↗ View details
A/1/3740
Astronomical bodies and related phrases: stars, galaxies, dwarfs, pulsars, etc
Zoom to ↗ View details
A/1/3847
Adjectives in math / linear algebra: " linear" etc.
Zoom to ↗ View details
A/1/1932
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2871
High energy physics
Zoom to ↗ View details
Cluster #308
A/0/84
This 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.
Zoom to ↗ View details
A/1/588
[Ultralow density cluster]
Zoom to ↗ View details
A/1/3711
Galactic astronomy (prepositions)
Zoom to ↗ View details
A/1/1950
" between"/" through"/similar prepositions in electromagnetics/quantum physics?
Zoom to ↗ View details
A/0/103
The feature fires in response to usage of the preposition "by" in contexts describing mathematical or scientific processes.
Zoom to ↗ View details
A/1/1555
Abstract verbs or endings, as might be done to equations or expressions, in math / theoretical physics? (" solve", " compute" etc.)
Zoom to ↗ View details
A/1/3429
" of" in [context]
Zoom to ↗ View details
Cluster #293
A/0/239
This feature seems to attend to mathematical expressions, formulas, and equations.
Zoom to ↗ View details
A/1/1626
" Note"
Zoom to ↗ View details
A/1/2979
Math nouns, predicting inline LaTeX
Zoom to ↗ View details
A/1/3090
This feature appears to activate when numbers or mathematical symbols appear in the context of equations, matrices, vectors, norms, spaces, parameters, eigenvalues, functions, etc.
Zoom to ↗ View details
A/1/3542
" theorem", " section", and other parts/units of math papers
Zoom to ↗ View details
A/1/3761
The feature appears to fire on symbols and terminology associated with mathematical expressions, equations, operators, and concepts.
Zoom to ↗ View details
A/1/3975
[Ultralow density cluster]
Zoom to ↗ View details
A/1/4054
Abstract nouns in mathematics after "the"
Zoom to ↗ View details
Cluster #295
A/0/318
This feature fires when it sees the word "the" in the token immediately preceding a technical term.
Zoom to ↗ View details
A/0/341
This 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.
Zoom to ↗ View details
A/1/512
" a" in the context of math, expected to be followed by a math word. E.g. "a subset" or "a Hilbert space"
Zoom to ↗ View details
A/1/1078
" the" in physics, esp electromagnetism?
Zoom to ↗ View details
A/1/1452
" the" in theory of machine learning / statistics
Zoom to ↗ View details
A/1/1638
" the" in math (advanced geometry? algebra?)
Zoom to ↗ View details
A/1/1652
" The" in [context]
Zoom to ↗ View details
A/1/1780
" the" in astrophysics?
Zoom to ↗ View details
A/1/3362
" the" in mathematics/physics/dynamical systems?
Zoom to ↗ View details
A/1/2183
Astronomy, especially solar-system
Zoom to ↗ View details
A/1/976
" a" in physics or other sciences contexts
Zoom to ↗ View details
A/1/3503
Famous mathematicians, predicting things named after them like " theorem"
Zoom to ↗ View details
A/1/1637
" for" in [context]
Zoom to ↗ View details
A/1/2395
" such" in [context]
Zoom to ↗ View details
A/1/1811
The feature fires when it sees scientific terms or concepts.
Zoom to ↗ View details
Cluster #304
A/0/357
This feature attends to words related to Coulomb interactions in a physics context.
Zoom to ↗ View details
A/1/128
Physics relating to motion and change: speed, rising/falling, developing/unstable/steady, etc
Zoom to ↗ View details
A/1/342
" electron"
Zoom to ↗ View details
A/1/1558
" of" in [context]
Zoom to ↗ View details
A/1/2878
Key actions, operands and logical constructs that are part of statistical analysis / modeling (prose)
Zoom to ↗ View details
A/1/3502
" with" in math (esp. geometry?)
Zoom to ↗ View details
Cluster #285
A/0/315
This feature attends to hyphens appearing in scientific writing.
Zoom to ↗ View details
A/1/216
Abstract nouns like "stability" in numerical systems, astrophysics, etc.
Zoom to ↗ View details
A/1/823
Graph theory (esp. adjectives)
Zoom to ↗ View details
A/1/3158
Nouns in math, esp. linear algebra ("matrix", "vector")
Zoom to ↗ View details
A/1/3435
Neuron networks/machine learning (prose)
Zoom to ↗ View details
A/1/1164
" function", esp. in biology
Zoom to ↗ View details
A/1/541
Physics-related noun of a particular flavor, e.g. "curvature", "velocity", "lattice", "transition"
Zoom to ↗ View details
A/1/341
This feature attends to scientific terminology related to dynamics, structures, gradients, perturbations, equations, behaviors, and parameters.
Zoom to ↗ View details
Cluster #277
A/0/16
This feature fires on word endings that describe geometric shapes and locations.
Zoom to ↗ View details
A/0/210
This feature fires when it sees technical terms and concepts commonly found in physics and engineering documents, particularly those related to optics and lasers.
Zoom to ↗ View details
A/1/65
" wave"
Zoom to ↗ View details
A/1/322
Physics (Optics)
Zoom to ↗ View details
A/1/2112
" horizontal"/" vertical" geometry?
Zoom to ↗ View details
A/1/3733
Graphics processing?
Zoom to ↗ View details
A/1/2134
" respectively"
Zoom to ↗ View details
A/1/118
Statistics (English text)
Zoom to ↗ View details
A/1/272
Abstract nouns esp related to paths in astronomy
Zoom to ↗ View details
Cluster #16
A/0/63
This feature fires when the word "according" is followed by the word "to".
Zoom to ↗ View details
A/1/890
" according" in [context]
Zoom to ↗ View details
A/1/1403
"According"
Zoom to ↗ View details
A/1/2064
" according" in [context]
Zoom to ↗ View details
A/1/1518
Graph/diagram-related nouns (curves, lines, squares, diagrams)
Zoom to ↗ View details
A/1/2744
Functions of mechanical systems / devices
Zoom to ↗ View details
A/1/800
" invention"
Zoom to ↗ View details
Cluster #276
A/0/31
The feature attends to text related to semiconductor transistors with respect to their components and fabrication.
Zoom to ↗ View details
A/0/129
This feature attends to technical language describing the physical structure or form of mechanical objects.
Zoom to ↗ View details
A/0/306
This feature attends to descriptions of rotating mechanical parts such as shafts, knobs, axles, and gears.
Zoom to ↗ View details
A/1/365
Prepositions and abstract verbs in mechanical engineering? metalworking?
Zoom to ↗ View details
A/1/521
Printing / recording / photographing
Zoom to ↗ View details
A/1/685
"relates to" (and similar phrases) in patent filings.
Zoom to ↗ View details
A/1/915
HVAC
Zoom to ↗ View details
A/1/1319
English nouns for mechanical components, predicting numbers (patents?)
Zoom to ↗ View details
A/1/1742
Semiconductors / Computer Manufacturing
Zoom to ↗ View details
A/1/2127
" crystal"
Zoom to ↗ View details
A/1/2485
English (Signal Processing)
Zoom to ↗ View details
A/1/3074
Material science phrases
Zoom to ↗ View details
A/1/3227
Electrical engineering
Zoom to ↗ View details
A/1/3792
Engineering devices (circuit, lamp, disk drive, instrument, device)
Zoom to ↗ View details
A/1/3901
Adjectives describing geometries of mechanical connections
Zoom to ↗ View details
A/1/3230
Concurrency, multiprocessing?
Zoom to ↗ View details
Cluster #259
A/0/81
This feature fires on text talking about technical concepts related to electronic communication, signal processing, and data transmission.
Zoom to ↗ View details
A/1/233
CPU architecture
Zoom to ↗ View details
A/1/307
Computer networking concepts
Zoom to ↗ View details
A/1/1841
Connector words in electronic storage/protocols
Zoom to ↗ View details
A/1/2066
Wireless networks
Zoom to ↗ View details
A/1/2096
" surface"
Zoom to ↗ View details
A/1/1166
Nouns or word fragments in biology, esp. animals or fungi
Zoom to ↗ View details
A/1/3631
" needle" and broadly similar nouns in sewing/crafting
Zoom to ↗ View details
A/1/2907
Abstract nouns in medicine and physics, often related to filtering
Zoom to ↗ View details
A/1/1301
Generic, inspecific terms for mechanical, biomechanical, and electronic components
Zoom to ↗ View details
A/1/1034
Words in scientific text, esp. atomic chemistry
Zoom to ↗ View details
A/1/3322
"aid" in "Chinese/Japanese Patent [Application] Laid-Open"
Zoom to ↗ View details
Cluster #272
A/0/254
The feature appears to fire on words and phrases related to chemical processes.
Zoom to ↗ View details
A/1/373
Excerpts from technical documents describing polymer chemistry processes and properties
Zoom to ↗ View details
A/1/2910
" with"/" containing" in chemistry
Zoom to ↗ View details
Cluster #260
A/1/241
"xe"
Zoom to ↗ View details
A/1/385
Phrases around chemical substances
Zoom to ↗ View details
A/1/1567
Chemical abbreviations
Zoom to ↗ View details
A/1/181
Chemistry " pH", " temperature", etc.
Zoom to ↗ View details
A/1/3287
" results"
Zoom to ↗ View details
A/1/3261
Statistics-related nouns or endings thereof, in "######" section titles
Zoom to ↗ View details
Cluster #312
A/0/218
The feature attends to words and phrases commonly found in sections describing the results of scientific papers.
Zoom to ↗ View details
A/1/3214
" study"
Zoom to ↗ View details
A/1/3253
" studies"
Zoom to ↗ View details
A/1/4076
" difference"
Zoom to ↗ View details
Cluster #311
A/0/83
The feature attends to words referring to different types of cells.
Zoom to ↗ View details
A/0/179
The feature attends to text describing figures and plots.
Zoom to ↗ View details
A/0/321
The feature attends to words and phrases related to cell death and injury.
Zoom to ↗ View details
A/1/159
This feature fires when it sees words related to chemical processes, such as polymerization, oxidation, precipitation, elution, etc.
Zoom to ↗ View details
A/1/336
The feature fires on tokens related to biological or medical samples.
Zoom to ↗ View details
A/1/997
Mostly " cells" and similar terms
Zoom to ↗ View details
A/1/1070
" Expression"/" expression"
Zoom to ↗ View details
A/1/1144
" age"
Zoom to ↗ View details
A/1/1434
" group"
Zoom to ↗ View details
A/1/1981
The feature fires when it sees words related to time periods or durations such as 'after', 'during', 'weeks', 'months', etc.
Zoom to ↗ View details
A/1/2510
This feature attends to words and phrases related to growth, proliferation, and increase.
Zoom to ↗ View details
A/1/2951
The feature fires when it sees words related to genes and genetic concepts.
Zoom to ↗ View details
A/1/3324
Some nouns esp. in medicine/science, but pretty varied
Zoom to ↗ View details
A/1/3332
Biomedical techniques
Zoom to ↗ View details
A/1/3470
Blood test results (as levels or concentrations)
Zoom to ↗ View details
A/1/3948
Abstract plural nouns in broadly technical/scientific contexts
Zoom to ↗ View details
A/1/4071
This feature attends to words related to research, diagnoses and treatments for diseases and disorders.
Zoom to ↗ View details
A/1/2321
[Ultralow density cluster]
Zoom to ↗ View details
A/1/1412
" population"
Zoom to ↗ View details
A/1/1840
The feature fires when it follows words that end with "s".
Zoom to ↗ View details
Cluster #286
A/0/49
This 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.
Zoom to ↗ View details
A/1/98
" count"/" phase" in various contexts
Zoom to ↗ View details
A/1/3772
Mechanical engineering performance
Zoom to ↗ View details
A/1/1086
" patients"
Zoom to ↗ View details
A/1/3750
" disease"
Zoom to ↗ View details
A/1/3131
Biology terms
Zoom to ↗ View details
A/1/907
" treatment"/" therapy" medical nouns
Zoom to ↗ View details
A/1/3760
Unit of time in days/months/years (but not minutes/hours)
Zoom to ↗ View details
A/1/1405
Demographic information stratified by age
Zoom to ↗ View details
A/1/1933
Synthetic biology
Zoom to ↗ View details
Cluster #317
A/0/41
This feature attends to wording related to perceptual and cognitive processing in scientific and research writing.
Zoom to ↗ View details
A/0/145
The feature attends to parenthetical information (e.g. numbers in parentheses).
Zoom to ↗ View details
A/1/482
"%" or verbs in stats from medical trials?
Zoom to ↗ View details
A/1/811
Medical imaging, esp. abbreviations like " CT"/" MR"
Zoom to ↗ View details
Cluster #169
A/0/108
The 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.
Zoom to ↗ View details
A/1/22in ____ condition" or similar phrases describing the state or condition of something. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2483) A/1/2604 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2604) A/1/2836 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2836) A/1/3058 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3058) A/1/3179 Various languages: " in" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3179) A/1/3616 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3616) A/1/3622 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3622) A/1/3721 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3721) A/1/3827 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3827) A/1/4049 " in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4049) A/1/47 " for" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-47) A/1/1224 'for' in biomedical text Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1224) A/1/1419 " order" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1419) Cluster #218 A/0/473 The feature attends to prepositional phrases containing the preposition "with". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-473) A/1/484 " from" in scientific studies Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-484) A/1/1046 " using" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1046) A/1/1679 " with" in science/photography? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1679) A/1/2462 " by" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2462) A/1/4028 " via" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4028) A/1/3483 " with" in technical contexts? web servers? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3483) A/1/2905 " with" in medicine / clinical trials? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2905) A/1/1031 " by" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1031) A/1/645 " of" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-645) Cluster #155 A/0/335 This feature fires when examples start with "In order to". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-335) A/1/627 Words shortly after sentence-opening "In" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-627) A/1/1451 " the" after "On" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1451) A/1/1511 " the" after "In" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1511) A/1/3315 " addition" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3315) A/1/2022 " from" in programming/software contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2022) A/1/2451 " considered" and broad synonyms in passive voice Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2451) A/1/2883 " clear" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2883) A/1/1245 Word fragments: passive verbs or some adjectives after "is"/word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1245) Cluster #303 A/1/381 Adjectives after "was"/"are" esp describing people Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-381) A/1/1392 The feature fires for words that have adjectival suffixes like -ful, -ous, -ible, -able, -al, etc. that indicate judgment or evaluation. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1392) A/1/2287 " crucial"/" important" and synonyms, esp. after "plays a/an" (predicting "role") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2287) A/1/3151 " unlikely"/" impossible" after "it is"/"it was" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3151) A/1/3916 Emotions, often " about" something Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3916) A/1/4077 " enough" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4077) Cluster #300 A/0/193 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-193) A/0/234 The feature attends to the verb "is defined as". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-234) A/1/165 " shown" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-165) A/1/195 " thrown" and similar action verbs likely to receive a direction/preposition Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-195) A/1/245 Past-tense verb about data in prose discussion of code Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-245) A/1/920 " used" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-920) A/1/1371 Adjectives describing known or established information, esp. " known", esp. after "is"/"are", in medicine Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1371) A/1/1414 Past-tense verb after "was" esp. in history/government Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1414) A/1/1632 Past-tense verbs/participles, passive voice, often describing conceptual actions like "defined"/"characterized" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1632) A/1/2020 " due" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2020) A/1/2143 " presented" and similar verbs in science Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2143) A/1/2290 'measured' etc in scientific papers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2290) A/1/2567 " showed" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2567) A/1/2578 This feature fires when it sees phrases describing a process of manufacturing, depositing, arranging or forming materials, substances or components. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2578) A/1/2911 " obtained" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2911) A/1/3494 This feature attends to words used to describe things being required, necessary, or expected. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3494) A/1/3513 " described" / " discussed" esp in legal contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3513) A/1/3523 " available" and some approximate synonyms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3523) A/1/3593 The feature fires when it encounters words ending in `ed`, `ly`, or `ing`, which are common past-tense, adverb, and progressive verb endings in English. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3593) A/1/3737 " compared" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3737) A/1/3780 " associated" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3780) A/1/3822 The feature fires on passive voice verbs and phrases indicating that something was done. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3822) A/1/3081 " resulted" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3081) Cluster #301 A/1/1796 Past-tense verbs like " interested"/" participated", for people in activities, predicting " in" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1796) A/1/2045 " arising" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2045) A/1/2507 " located" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2507) A/1/3448 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3448) A/1/3708 " exist"/" occur" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3708) A/1/1023 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1023) Cluster #270 A/0/12 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-12) A/1/894 Adverbs that indicate aspects of time Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-894) A/1/2329 Adjectives after "has"/"had", often size-related metaphorically describing results or plans Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2329) A/1/3962 Words predicting me / myself / I ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3962) A/1/2314 "Thank" (maybe religious?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2314) A/1/3871 " me" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3871) A/1/13 " thanks" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-13) A/1/4060 Pronouns mainly him and them Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4060) Cluster #164 A/0/198 The feature fires when it sees the word "something" and words related to the general concept of something. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-198) A/1/458 " itself" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-458) A/1/1446 " here" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1446) A/1/1607 " nothing" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1607) A/1/2021 " anyone"/" someone"/" somebody" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2021) A/1/2450 " something" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2450) A/1/3709 " Everything"/" everything" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3709) A/1/3820 "somewhere else" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3820) Cluster #266 A/0/237 This feature fires on the verb "do" and its variations such as "did", "done", and "doing." Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-237) A/1/302 "please", sometimes capitalized or with leading space Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-302) A/1/629 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-629) A/1/3009 " well" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3009) A/1/3206 " let" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3206) A/1/3317 " doing" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3317) A/1/3895 " do" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3895) A/1/4001 " need" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4001) A/0/492 This feature fires when the token "me" appears. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-492) A/1/1312 Spiritual, self-help style psychology/charity text, of the sort that would be titled 'Giving is the Key to Happiness' Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1312) A/0/102 This feature fires when it detects words related to gaining attention or awareness. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-102) Cluster #279 A/1/640 This feature fires when it sees words about physical interactions such as hands, arms, fingers, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-640) A/1/2320 Nouns, esp. abstract, after "make" or word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2320) A/1/2707 Pronouns after "make" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2707) A/1/3013 Indirect object pronouns Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3013) A/1/3433 " care"/" custody"/etc. after "take" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3433) A/1/3890 " together" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3890) A/1/417 Object pronouns after "tell" or word forms (predicting " why"/" how"/etc.) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-417) Cluster #282 A/1/1131 " able" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1131) A/1/1241 Past participles after "have"/"has"/"had", usually transitive and predicting object pronouns Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1241) A/1/2166 " unable"/" fail" esp after "was"/"were"/etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2166) A/1/3224 The feature attends to sentences that describe someone's feelings or emotions. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3224) A/0/330 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-330) A/1/1408 " want" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1408) A/1/2275 " trying" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2275) Cluster #306 A/0/65 This feature is activated when it detects an instance of the phrase "going to" or related variations (such as "gonna", "go ahead", "go back", etc.). Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-65) A/1/1422 " bet" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1422) A/1/1613 " going" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1613) A/1/2154 " try" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2154) A/1/3874 " go" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3874) A/1/3994 " went" (some " gone"/" goes"/... Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3994) A/1/3 "worth" single-token feature, esp. after "is" or word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3) A/1/2234 " said" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2234) Cluster #294 A/0/138 The feature seems to fire when expressions of thinking such as "think", "figured", "believe", etc. are used. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-138) A/0/480 The feature fires for phrases related to knowing or understanding. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-480) A/1/585 " believe" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-585) A/1/775 " forget" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-775) A/1/1080 Present-tense verb after / near "you"/"You" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1080) A/1/1234 " say" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1234) A/1/1334 " told"/" tells"/word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1334) A/1/1381 " know" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1381) A/1/1651 " sure" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1651) A/1/2386 " hope" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2386) A/1/2420 " mean" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2420) A/1/2482 " means" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2482) A/1/2582 " show" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2582) A/1/2630 " see" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2630) A/1/2751 " heard" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2751) A/1/2797 " think" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2797) A/1/3828 " thinking"/" thought" word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3828) A/1/3929 Verbs after "I", esp. present-tense Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3929) A/0/235 The feature fires when it sees dialogue and related tags such as "he/she said", "she replied", "he told me", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-235) A/1/1119 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1119) Cluster #322 A/0/333 This feature seems to fire on the base verb "take" in a variety of grammatical contexts and tenses. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-333) A/0/468 This feature fires when it sees verbs related to giving or making (give, makes, made, keeping, etc). Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-468) A/0/497 This feature seems to fire when a form of the verb "get" is present. Examples include "getting", "got", "gotten", "get", "gets". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-497) A/1/500 " keep" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-500) A/1/554 " take" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-554) A/1/1666 "make" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1666) A/1/1747 " taken" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1747) A/1/1880 " spent"/" spend"/etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1880) A/1/2658 This feature fires when it sees verb forms related to the word "bring". The feature attends to words like bring, brought, bringing, and brings. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2658) A/1/3495 " put" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3495) A/1/3628 " give" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3628) A/1/3984 " lose" word forms, or rarely " gain" word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3984) A/1/4025 " find" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4025) A/1/251 " wanted" and other verbs of volition Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-251) Cluster #323 A/0/381 This feature fires when it sees verbs and adjectives related to visual perception and appearance such as "look", "seem", "stare", "glow", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-381) A/1/2225 Active/physical verbs, performed by personal pronouns Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2225) A/1/3264 " look" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3264) A/1/3888 " speak" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3888) A/1/1075 " come" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1075) A/1/1943 This feature fires on the lemma "get" in various forms (gotten, getting, etc). Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1943) A/1/2933 "looked" and "looking" as verbs to describe the act of visually directing one's gaze, as well as more figurative usages indicating scrutiny, seeking. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2933) A/1/1615 " killed" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1615) A/0/58 This feature fires when it sees words indicating that someone or something has arrived from elsewhere. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-58) A/1/1843 The feature activates when it sees words related to beginning, such as began, started, finished, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1843) Cluster #331 A/0/440 This feature fires when the word "led" appears. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-440) A/1/648 The feature attends to verbs that indicate the role something plays, such as serves, referred, attributed, known, appointed, joined, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-648) A/1/809 " met" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-809) A/1/3816 " receive" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3816) A/1/4075 " followed" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4075) A/1/4046 " directed" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4046) A/1/3446 " led" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3446) A/1/3262 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3262) A/1/1957 " return" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1957) A/1/3511 " decided" / " agreed" etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3511) A/1/2722 " happen" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2722) Cluster #333 A/0/60 This feature seems to fire on verbs related to physical actions and movements like pull, set, carry, move, cut, burn, freeze, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-60) A/1/351 " drink", " eat", and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-351) A/1/1003 " break" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1003) A/1/1104 " turned", turns, turning - token Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1104) A/1/1156 " set" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1156) A/1/1838 " carried" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1838) A/1/2913 " pick" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2913) A/1/3195 " cut"/" pull"/" knock" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3195) A/1/3536 " play" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3536) A/1/3681 " follow" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3681) A/1/3713 " buy" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3713) A/1/2516 " talking" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2516) A/1/174 " read" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-174) A/1/215 " wait"/" waiting" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-215) A/1/2374 " fall" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2374) Cluster #332 A/0/463 This feature fires when a sentence contains "out" as part of a directional phrasal verb. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-463) A/1/86 Primarily " died", few other past-tense verbs for major life events Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-86) A/1/347 " grow"/" grown" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-347) A/1/958 " up" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-958) A/1/1216 " off" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1216) A/1/2315 " out" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2315) A/1/2403 " out" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2403) A/1/3071 " start" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3071) A/1/3204 " move" and word forms, of people/things, predicting explanations of how it moved Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3204) A/1/3267 " back" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3267) A/1/3417 " away" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3417) A/1/3438 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3438) A/1/3321 " called" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3321) A/1/2245 " interact" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2245) A/1/1027 " beginning"/" starting" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1027) Cluster #329 A/1/1901 " living"/" staying" and forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1901) A/1/3375 " stand" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3375) A/1/3612 " lie" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3612) A/1/1636 This feature seems to fire on past tense verbs, especially ones ending in -ed. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1636) A/1/1822 Past-tense verbs for physical actions performed by people Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1822) A/1/3139 Word endings with "-ing" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3139) A/1/552 " left" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-552) Cluster #327 A/0/32 This feature seems to attend to verbs that indicate some sort of change, such as reassign, switch, upgrade, enact, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-32) A/1/765 " create"/" develop" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-765) A/1/3370 Abstract political actions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3370) A/1/3945 The feature seems to fire on verbs related to concealing or revealing information. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3945) A/1/4026 " help" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4026) A/1/526 " agree" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-526) A/1/1728 " learn" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1728) Cluster #330 A/0/263 This feature fires when it sees words related to establishing or determining something. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-263) A/1/691 " provide" and word forms, usually in an abstract sense Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-691) A/1/1721 " suffer" and word forms, sometimes other verbs related to injuries Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1721) A/1/2119 " seek" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2119) A/1/2375 This 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". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2375) A/0/496 This feature fires when it sees present participle verb forms (-ing) as well as some similar sounding endings like -ington, -ation, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-496) A/1/1306 'choose' token, literal meaning of choice Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1306) A/1/2999 " love" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2999) A/1/2530 " ensure" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2530) Cluster #176 A/0/2 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-2) A/1/255 " based" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-255) A/1/1133 " rely"/" focus" word forms etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1133) A/1/1983 " depending" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1983) A/1/2639 Gerunds related to cleaning/physical maintenance Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2639) A/1/3953 " inspire"/" enjoy"/positive language Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3953) A/1/2345 " opening" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2345) A/1/1024 " determine"/" ascertain" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1024) Cluster #287 A/0/434 This feature is activated by examples describing the act of doing work or a job. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-434) A/1/617 " act"/" acts"? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-617) A/1/785 " check", esp. " check out" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-785) A/1/2746 " works" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2746) A/1/2787 "change" as a process of transition, development in living organisms, natural systems, social dynamics and human institutions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2787) A/1/2825 " working" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2825) A/1/3491 " work" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3491) A/1/369 " pay" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-369) A/1/3793 " sat" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3793) A/0/56 This feature attends to phrases and words related to contacting and getting in touch such as "Contact us", "Call today", "Pick up" etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-56) A/1/3255 " Google"/"Google" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3255) A/0/245 The feature fires when it sees case law citations such as "State v. DefendantName" and legal language such as "appellant" and "jurisdictional status". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-245) A/1/1352 Tokens in the middle or at the end of names in legal contexts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1352) A/1/1581 Capitalized start-of-sentence imperatives/instructions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1581) A/1/880 Title-case, mainly business/marketing-related words Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-880) A/1/38 Party in a court case of "A vs B" formatting Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-38) Cluster #234 A/1/483 " Supreme" in "Supreme Court", often following a U.S. state Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-483) A/1/3102 Title Case words in names of laws, legal agreements, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3102) A/1/3933 " Act" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3933) Cluster #236 A/1/1601 " United" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1601) A/1/2494 Capitalized compass directions esp. " North" predicting plausible location names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2494) A/1/2719 Directional adverb in geography Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2719) A/1/3677 " News" and other names of news organizations Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3677) A/1/686 Title Case Nouns esp. undelimited titles in sentences Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-686) Cluster #237 A/1/452 " Ever" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-452) A/1/563 " New" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-563) A/1/1160 This feature fires when it sees black (the color or race) in text. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1160) A/1/1300 " High" in high school names, e.g. "Riverdale High School"; or rarely " Middle" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1300) A/1/1799 " Big" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1799) A/1/2522 " Golden"/"Golden" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2522) A/1/3476 " Good"/"Good" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3476) Cluster #252 A/0/461 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-461) A/0/485 The feature fires when it sees proper nouns such as names of people, places, organizations, acronyms, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-485) A/0/508 This feature attends to proper nouns that contain uppercase words or abbreviations in the middle of a sentence. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-508) A/1/80 Paper acknowledgements, esp. names of institutions therein Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-80) A/1/176 Title Case (News Headline) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-176) A/1/308 Title Case formal model names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-308) A/1/419 This feature fires on words and phrases related to legal cases, such as "Case:", "Claim No.", "Case No.", "Court", "Order", "Form", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-419) A/1/701 Capitalized scientific publications/institutions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-701) A/1/1172 Title Case words for abstract units of data or relating to finance/payments Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1172) A/1/1413 Capitalized terms for organizations Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1413) A/1/1587 " Health" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1587) A/1/1703 " Law" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1703) A/1/2046 Title Case Medical terms, e.g. in titles of papers/studies Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2046) A/1/2075 " War", esp. in "World War", predicting " II" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2075) A/1/2126 " Down" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2126) A/1/2232 Capitalized, esp. last, words in formal names of organizations/awards/etc. but before the "kind of thing" it is (e.g. not " Committee" or " Association") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2232) A/1/2300 " Time" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2300) A/1/2772 Title-case words, esp. " Dream", esp. used in a "compound word" name like "Dreamcast" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2772) A/1/3166 " National" in Title Case organization names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3166) A/1/3307 Academic Institution Name Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3307) A/1/4066 " World" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4066) Cluster #251 A/0/99 This feature fires when proper nouns are present, especially locations like cities and states. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-99) A/1/293 Names of roads and streets Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-293) A/1/306 Countries or continents, capitalized Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-306) A/1/687 Places with bipartite names (_ Creek, _ Falls) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-687) A/1/2566 " City" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2566) A/1/3447 Major American city names, predicting locations/things in them, e.g. " City" or " County" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3447) A/1/4018 " York" in "New York", predicting things in New York, e.g. " City" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4018) A/1/1985 Music/Dance/Broadway/Movies named entities Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1985) A/1/1752 Ends of Title Case company/brand names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1752) A/1/3869 " Part" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3869) A/1/2904 " General" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2904) Cluster #229 A/0/337 This feature fires on italics or book/article titles. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-337) A/1/1665 Title Case words / word ends in book titles Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1665) A/1/2851 Underscore-delimited names/words Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2851) A/1/3861 Capitalized royal/religious titles (" Lord", " "King", etc.) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3861) A/1/2070 Mostly ends of proper nouns / names of people, predicting past-tense verbs they might perform Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2070) A/0/297 This feature attends to names ending in a, ie, or y. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-297) A/1/3456 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3456) A/1/3936 " Lock" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3936) A/1/2705 " Out" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2705) A/1/2460 Title Case location names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2460) A/1/2312 The feature fires when words contain "wood". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2312) A/1/346 Common first names, predicting past-tense verbs Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-346) A/1/394 " Inc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-394) Cluster #357 A/0/445 This feature fires on English names in text, particularly last names. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-445) A/1/123 Names of authors/locations/publishers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-123) A/1/646 This feature fires when it sees words containing the substring "ley". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-646) A/1/672 Common English last names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-672) A/1/1084 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1084) A/1/1190 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1190) A/1/2293 The feature seems to fire on Spanish surnames ending in vowels such as a, e, i, o. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2293) A/1/2703 Ends of names, particularly people, in English Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2703) A/1/2918 Roman/Greek proper names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2918) A/1/2973 Names of people or final tokens thereof Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2973) A/1/3098 Last token of surname, in legal descriptions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3098) A/1/3240 " Trump" (Donald Trump) / some other U.S. presidents Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3240) A/1/3664 English, esp. British, place names, predicting "shire" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3664) A/1/3818 Tokens in the middle or at the end of a name that looks like a location, predicting e.g. " County" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3818) A/1/3060 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3060) Cluster #354 A/1/17 "er" and "ter", usually near ends of words Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-17) A/1/101 "er" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-101) A/1/525 Ending of a person's name Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-525) A/1/1240 This feature fires for words whose final syllable sounds like "ore". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1240) A/1/2041 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2041) A/1/2123 The feature attends to tokens that end or have a "y" sound. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2123) Cluster #364 A/0/370 This feature attends to male first names ending with "n" (and variations) such as Kevin, Steven, Ken, John, Glenn, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-370) A/1/124 Typical English honorifics "Mr"/"Ms"/"Mrs"/"Miss" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-124) A/1/1244 Common English first names, esp. male, predicting last names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1244) A/1/2697 " Dr" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2697) A/1/891 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-891) Cluster #363 A/0/314 The feature fires when there are names or words ending with "in". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-314) A/1/1592 This feature attends to foreign names and words, especially German ones. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1592) A/1/2036 Parts of names, esp. endings and esp. people's last names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2036) A/1/3717 This 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". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3717) Cluster #365 A/1/1456 " Tim" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1456) A/1/3250 " Christ"/"Christ" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3250) A/1/3970 " Ben" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3970) A/1/1969 Tokens in middle or end of names, esp. location names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1969) A/1/1805 Ends of names or usernames, and "gate" in "enronXgate" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1805) A/1/2712 "e" in the end of last names of people Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2712) A/1/1056 Conjunctions/prepositions near Indian locations/government Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1056) A/1/231 Transliterated Arabic names of people? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-231) A/1/1093 Beginnings of names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1093) Cluster #362 A/0/457 This feature fires on words and names ending in "a". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-457) A/1/1053 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1053) A/1/1827 This feature appears to activate on syllables and sounds at the end of words like "ah", "uh", and "ih". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1827) A/1/1926 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1926) A/1/2136 This feature attends to Japanese, Korean, Tamil and Hindi names and words. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2136) A/1/3256 Mid- or end-word "as" or "is" fragments Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3256) A/1/3501 Mostly "la" but various non-English word/name endings ending in "a" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3501) Cluster #361 A/0/260 This feature attends to proper nouns ending in "el", "en" or "er". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-260) A/1/839 "in" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-839) A/1/939 Parts of a large variety of various names (Korean, Japanese, Persian, Yoruba...) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-939) A/1/1305 Single-lowercase-letter tokens, e.g. "b"/"l"/"n"/"m"/"k", inside capitalized words and names with unusual spellings Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1305) A/1/3842 This feature fires when it sees names or words ending in -en. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3842) A/1/1502 People's names, particularly Spanish/Latin-American Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1502) A/1/1848 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1848) Cluster #390 A/0/481 This feature attends to the letters "ul" surrounded by consonants. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-481) A/1/280 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-280) A/1/411 Tokens before German name suffixes like -berg, -stein, -reich, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-411) A/1/856 ~2-letter tokens after the start of a title-cased name, esp. "ul" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-856) A/1/865 Token after " Ca", either name continuation or number/punctuation for Calcium Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-865) A/1/1354 "ac" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1354) A/1/1417 Tokens in a capitalized word esp. after "Con" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1417) A/1/1454 Tokens in a capitalized word esp. after "Sch" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1454) A/1/1687 " McG" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1687) A/1/1962 Tokens in names, esp. " Lind" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1962) A/1/2196 This feature fires when it sees words containing double consonant letters and diphthongs in Spanish. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2196) A/1/2418 Space + capital letter after a first name, either a middle initial or starting a last name Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2418) A/1/3283 The feature appears to fire on proper nouns, especially names, that start with hard "k" or "c" sounds. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3283) A/1/3488 Mid-title-case-name tokens esp. "it"/"ac" after " Cap" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3488) A/1/3926 " Jr" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3926) A/1/1602 "ist" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1602) A/1/904 "mb" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-904) Cluster #35 A/0/189 This feature fires for two or three initials followed by a period, which is a common way of citing author's names in academic papers. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-189) A/1/509 " Rev" in in citation-like things, most prominently "Phys. Rev." Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-509) A/1/1291 Advanced math book titles Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1291) A/1/2369 First initial of author in citations (usually S) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2369) A/1/3075 Capital initials and initial capital letters of names, esp. in bibliographies, e.g. "S" in "S. S. Smith" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3075) A/1/3839 Middles and ends of European last names, esp. in bibliographies Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3839) A/1/3694 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3694) Cluster #353 A/1/1153 German: ch? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1153) A/1/2002 "av"/"ov" and other mid-word/name tokens with "v" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2002) A/1/2947 "ik" in last/first names from Slavic or Eastern European origins Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2947) A/1/3954 The feature fires when it sees words with the "ang" phoneme in them, such as bang, rang, fang, anguish, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3954) A/1/965 Sanskrit, Arabic, and Indic: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-965) A/1/3960 " Phys" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3960) Cluster #351 A/0/230 The feature seems to attend to foreign names and terms, primarily names from South Asian languages such as Sanskrit. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-230) A/1/1028 Japanese: traditional names written in Romaji Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1028) A/1/1448 Pinyin-transliterated Chinese names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1448) A/1/1579 Russian: Names connected to the country, not necessarily ethnically Russian Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1579) A/1/3639 English transliterations of Indian names? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3639) A/1/683 Word component "ost" or "st" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-683) A/1/3619 This feature attends to the suffix -ill and names ending in -ill. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3619) A/1/1447 Space + single capital letter tokens, typically in title case Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1447) A/1/1919 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1919) A/1/2291 "am" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2291) A/1/2621 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2621) A/1/384 Spanish first names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-384) A/1/2786 " Gi" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2786) A/1/4048 European languages: "az"/"iaz" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4048) A/1/778 " Man" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-778) Cluster #391 A/0/345 This feature activates for words containing the sound "cap" (cap, capacitance, capacitors, etc). Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-345) A/1/478 " Mon"/"Mon" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-478) A/1/1296 " Bon" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1296) Cluster #389 A/1/1836 " Santa" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1836) A/1/2304 " Saint" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2304) A/1/2660 " San"/"San" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2660) A/0/383 This feature appears to fire on names and places, particularly with Spanish origin such as Benigno Aquino, Cristo Yacente, María de los Milagros, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-383) A/1/2900 " St"/"St" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2900) A/1/1732 "enn" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1732) A/1/3265 "und" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3265) A/0/236 This feature fires for the prefix abbreviation 'St.', most commonly as part of location names such as cities and states. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-236) A/1/3638 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3638) A/1/1886 " Math" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1886) A/1/210 " Ann" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-210) A/1/1803 Pinyin-transliterated Chinese names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1803) A/1/2260 This feature attends to words that end in -ur. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2260) A/1/1329 " Bow" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1329) A/1/1142 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1142) A/1/2818 " Van" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2818) A/1/3108 " Cast" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3108) A/1/2479 " Fran" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2479) A/1/4047 "ar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4047) A/1/4038 " Com" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4038) A/1/1533 Mid-word "og" or other fragments ending with "g" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1533) A/1/909 " Am" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-909) A/1/2160 " Ok" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2160) A/1/2538 " Met" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2538) A/1/1212 " Jo" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1212) Cluster #392 A/0/223 This feature fires when a word includes the syllable "rog". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-223) A/1/58 " Hen", handful of " Ken"/" Len"/name starts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-58) A/1/431 " Med" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-431) A/1/594 " Ang" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-594) A/1/696 " Ser"/"Ser" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-696) A/1/790 " Mac"/"Mac" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-790) A/1/836 " Mar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-836) A/1/924 "Mont" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-924) A/1/1180 " Mur" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1180) A/1/1384 " Mad" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1384) A/1/1519 Beginnings of proper nouns esp. " Em"/" Ham" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1519) A/1/1853 " Er" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1853) A/1/1890 The feature attends to words with /məl/, /i/, and /ɑː/ sounds (such as Mal-, I-, A-) in the middle or end of words. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1890) A/1/1948 " Brit" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1948) A/1/1984 " Mc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1984) A/1/2157 " Mer" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2157) A/1/2527 " Mor" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2527) A/1/2627 " Har" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2627) A/1/2644 " Luc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2644) A/1/2809 " Rem" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2809) A/1/2981 " Bur" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2981) A/1/3185 " Jan" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3185) A/1/3424 " Mah" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3424) A/1/3527 " Mag" etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3527) A/1/3562 " Pet" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3562) A/1/3673 " Ed" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3673) A/1/3702 " Cor" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3702) A/1/3794 " Pat" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3794) A/1/3802 " Mart" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3802) A/1/2003 " Az" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2003) A/1/2261 " Bel" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2261) A/1/1258 " Hor" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1258) A/1/3452 " Sim" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3452) A/1/2052 " Sign" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2052) A/1/2732 "it" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2732) Cluster #393 A/0/382 The feature attends to words ending in "-al", especially when it is part of a proper noun. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-382) A/0/471 This feature fires on words that contain the syllable "terra" such as territory, terracing, Terran, and terraced. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-471) A/1/147 " sal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-147) A/1/153 " Wil" in names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-153) A/1/352 " Car" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-352) A/1/548 " Col" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-548) A/1/566 "ou" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-566) A/1/984 " Ter"/"Ter" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-984) A/1/1219 " Sac" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1219) A/1/1536 "mar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1536) A/1/1621 " ger" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1621) A/1/1630 " Sand" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1630) A/1/1730 " Par" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1730) A/1/1733 " Sy" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1733) A/1/1896 " El" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1896) A/1/1925 " Ban"/"Dun"/" Dun", esp. beginning names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1925) A/1/1952 " Kar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1952) A/1/2282 " Bal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2282) A/1/2340 " ar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2340) A/1/2352 " Anal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2352) A/1/2517 " Gal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2517) A/1/2573 " Hel" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2573) A/1/2704 "ij" in "ijerph" citations ("@B1-ijerph-...") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2704) A/1/2845 " Pol" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2845) A/1/2875 " Al" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2875) A/1/3047 " Cal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3047) A/1/3062 " Grand"/"Grand" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3062) A/1/3246 " Ber" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3246) A/1/3331 " Far" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3331) A/1/3404 " Vol" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3404) A/1/3521 " Ly" and similar usually capitalized word beginnings Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3521) A/1/3522 " Ar" and similar usually capitalized word beginnings Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3522) A/1/3800 " Sher" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3800) A/1/3876 " Del"/"Del" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3876) A/1/3999 " Wal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3999) A/1/4065 " Val" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4065) A/1/2150 "leg" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2150) A/1/2090 " bar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2090) A/1/942 Mid-capitalized-word tokens ending in "t", sometimes "ct", predicting "orial" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-942) A/1/1532 " Port"/"Port" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1532) A/1/2515 Ver Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2515) A/1/3274 " Mat" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3274) Cluster #382 A/1/2940 " Dam" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2940) A/1/2953 " Mot" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2953) A/1/3126 " Sav" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3126) A/1/1439 " Gen" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1439) Cluster #381 A/1/1037 " Ant" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1037) A/1/1361 " Ind" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1361) A/1/2008 " ed" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2008) A/1/2547 Mostly " Ent", some "Jac"/"Vent" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2547) A/1/3295 " Vis" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3295) A/1/502 "os" ending a word (often Spanish word or name) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-502) Cluster #371 A/1/595 " Res" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-595) A/1/888 " Den" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-888) A/1/1483 " Nat" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1483) A/1/2197 " pul" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2197) A/1/3725 " Mus"/"Mus" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3725) A/1/3427 Word fragments predicting "let" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3427) A/1/737 " Pan"/" pan"/"Pan" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-737) A/1/1960 " Pal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1960) A/1/3396 " Fre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3396) A/1/1508 " Sil" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1508) A/1/763 " Appro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-763) A/1/276 " sol" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-276) A/1/3200 " Di" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3200) A/1/2142 " Re" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2142) A/1/1550 "co" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1550) A/1/3379 " La" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3379) Cluster #384 A/0/365 This feature attends to names or words starting with 'Pe' or 'Se'. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-365) A/1/203 " Pe" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-203) A/1/274 " Co" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-274) A/1/425 " Me"/"Me" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-425) A/1/893 " Be" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-893) A/1/1401 " Bre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1401) A/1/1487 " Se"/"Se" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1487) A/1/1566 " Le" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1566) A/1/1794 " Ke" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1794) A/1/2039 "ste" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2039) A/1/2159 " Che" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2159) A/1/3109 " Ste" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3109) A/1/3734 " Ne" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3734) A/1/2508 This feature attends to words ending in "o" or containing "cho". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2508) A/1/117 " De" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-117) Cluster #385 A/0/176 This feature attends to tokens that start with Sa or Ba. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-176) A/1/722 The feature attends to words that start with "Ro". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-722) A/1/1591 " Ha" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1591) A/1/1628 " Bo" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1628) A/1/2258 " Ca" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2258) A/1/2343 " Sa" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2343) A/1/2426 "ri" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2426) A/1/3083 " po" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3083) A/1/3458 " ca" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3458) A/1/3658 " Lo" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3658) A/1/590 " Ki" and other space, capitalized-consonant, vowel tokens Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-590) Cluster #383 A/0/361 The feature seems to fire when words start with "tra". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-361) A/1/200 " Fra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-200) A/1/423 "ra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-423) A/1/667 " Tre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-667) A/1/1276 " fe" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1276) A/1/1643 " Bro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1643) A/1/1685 "Tro"/" Tro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1685) A/1/1855 " pri" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1855) A/1/2977 " Cro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2977) A/1/3667 " Bra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3667) A/1/3705 " Tra"/" tra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3705) A/1/3798 " Dra"/" dra"/"Dra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3798) A/1/3803 "ro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3803) A/1/3819 "du" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3819) A/1/3832 " Gra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3832) A/0/62 This feature seems to fire on co- words, pairs that share a prefix such as co-author, co-founder, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-62) A/1/1387 " co" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1387) A/1/2598 " tri" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2598) A/1/3443 " Ju"/"Ju" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3443) A/1/1248 " fre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1248) A/1/4003 " ri" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4003) A/1/2968 " cal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2968) A/1/4093 Mid-word "iqu", start-of-word "nu", and other "-u" tokens Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4093) A/0/506 The feature attends to occurrences of "et al" in citations or lists of authors. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-506) A/1/2027 " et" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2027) Cluster #335 A/1/636 Start-of-word "vari"/"propri", predicting "ety" and forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-636) A/1/744 " fore" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-744) A/1/3690 " dise" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3690) A/1/3489 " al" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3489) A/0/131 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-131) A/1/2723 " hal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2723) A/1/1333 " Cy"/"Cy" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1333) A/1/439 " Ad" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-439) A/1/3683 " per" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3683) A/1/304 "Per", as a word prefix Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-304) A/1/2085 " en", generally not Spanish Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2085) A/1/2863 " Sur" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2863) A/1/285 " En" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-285) A/1/132 " Ge" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-132) A/1/270 French or German: sur Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-270) A/1/412 " ap" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-412) A/1/2082 " Pro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2082) A/1/53 " Impro"/" impro" with some " Retrie" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-53) A/1/3154 " hy"/"Hy" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3154) Cluster #315 A/0/369 This 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". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-369) A/1/54 "con" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-54) A/1/261 " Con" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-261) A/1/689 " ex" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-689) A/1/1460 " con" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1460) A/1/2975 " Ex" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2975) A/1/3111 " trans" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3111) Cluster #324 A/0/54 This feature attends to prefixes dis- and in-, which are often used to negate words or express the opposite meaning. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-54) A/0/332 This feature attends to words starting with "re" that are often prefixes or suffixes. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-332) A/0/410 This feature seems to attend to words with the "un-" prefix, including misspellings or errors such as "untrained" or "uncertificated". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-410) A/1/998 " Un"/"Un", esp. right after EOT or in news/politics Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-998) A/1/2047 " mis" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2047) A/1/2646 " re" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2646) A/1/3067 " pro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3067) A/1/3451 " Dis" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3451) A/1/3998 " un" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3998) A/1/1322 "Pre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1322) A/1/2149 "Tran"/" Tran", predicting "script" and forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2149) A/1/1433 " dis" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1433) A/1/1368 " Des" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1368) Cluster #325 A/0/422 This feature fires when it sees prefixes like: pre-, non-, intra-, inter-. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-422) A/1/1055 " inter" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1055) A/1/1171 " sub" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1171) A/1/1515 " pre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1515) A/1/1660 " micro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1660) A/1/2203 " bio" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2203) A/1/2412 " anti" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2412) A/1/2445 " hyper" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2445) A/1/2633 " non" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2633) A/1/2677 " intra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2677) A/1/3389 " super" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3389) A/1/3855 " Sub" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3855) A/1/4020 " multi" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4020) A/1/3479 " mid" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3479) Cluster #372 A/0/50 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-50) A/1/1092 " Es"/" es" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1092) A/1/3344 " Ass" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3344) A/1/3545 Start-of-word "Sus"/" Sus" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3545) A/1/3620 " tele" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3620) A/1/2461 " neuro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2461) A/1/443 " electro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-443) A/1/464 " Gu" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-464) A/1/294 " mal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-294) A/1/746 " Nut" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-746) Cluster #374 A/1/551 Mid-word token like "ib", containing vowel followed by "b" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-551) A/1/1690 " Lib" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1690) A/1/1719 " Ab" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1719) A/1/3892 " cor" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3892) A/1/3769 " McK" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3769) A/1/3570 " res" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3570) A/1/1196 " Uns"/"Uns" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1196) A/1/3745 " Fer" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3745) A/1/2425 Mostly " Benef"/" Jud" predicting "iciary" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2425) A/1/2959 " ammon" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2959) A/1/989 " ep" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-989) A/1/3915 " ant" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3915) A/1/1372 "colog" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1372) A/1/2829 " Art"/" art" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2829) A/1/1588 " min" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1588) A/1/2649 "ob" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2649) A/0/338 This feature attends to the string "ab" occurring anywhere in a word, regardless of case. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-338) A/1/1942 " ab" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1942) A/1/3260 " ob" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3260) A/1/3599 " imm" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3599) A/1/1481 "ap" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1481) A/1/3492 " Emp"/"Emp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3492) Cluster #380 A/0/478 This feature attends to words related to admissions such as "admission", "admissions", "admissible", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-478) A/1/1605 " Bot" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1605) A/1/1620 " Cap" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1620) A/1/1688 " rect" and "scar", predicting "ify" and forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1688) A/1/1708 " mod" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1708) A/1/1907 " gall" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1907) A/1/2194 " Log"/" log" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2194) A/1/2249 " Net"/" net" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2249) A/1/2803 " Reg" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2803) A/1/3087 " col" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3087) A/1/3132 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3132) A/1/3198 " ad" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3198) A/1/3465 " Bet" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3465) A/1/2763 " cur" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2763) A/1/3496 " reg" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3496) A/1/344 " amp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-344) A/1/761 " App" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-761) Cluster #396 A/0/279 The feature seems to fire when it sees words related to computing and computers, such as "computed", "computationally", "computer", etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-279) A/1/115 " ag" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-115) A/1/221 " inf" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-221) A/1/258 " ext" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-258) A/1/750 " rep" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-750) A/1/771 " amb" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-771) A/1/1539 " imp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1539) A/1/1650 " ref" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1650) A/1/1835 " emb" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1835) A/1/2081 " exp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2081) A/1/2700 " Comp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2700) A/1/3137 " conf" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3137) A/1/3225 " Dep" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3225) A/1/3279 " Sup" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3279) A/1/3328 " Imp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3328) A/1/3850 " Ref" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3850) A/1/1202 " Ev" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1202) A/1/1673 " cent" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1673) A/0/77 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-77) A/1/2596 "Rep", esp. right after EOT and short for (U.S.) Representative Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2596) Cluster #388 A/0/423 The feature seems to fire for tokens containing the word "hop" (and related words like "hip"), specifically in the context of hip-hop music. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-423) A/1/319 " bur" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-319) A/1/599 " wrink" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-599) A/1/601 " scal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-601) A/1/1091 " hel" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1091) A/1/1360 Mostly " pot" with some " bull" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1360) A/1/1913 " bount"/" mas"/" bal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1913) A/1/3356 " bug" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3356) A/1/3512 " pop" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3512) Cluster #397 A/1/1554 " Int" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1554) A/1/1571 " def" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1571) A/1/2339 " Def" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2339) Cluster #386 A/1/208 " compress", "uggest", and similar penultimate word tokens, predicting suffixes like "ive"/"or" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-208) A/1/496 " cult" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-496) A/1/2458 " apost" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2458) A/1/2828 Mostly " gam" ("gamers"/"gamblers"/"gamification") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2828) A/1/883 " Cont"/"Cont" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-883) A/1/1906 " cat" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1906) Cluster #398 A/1/808 " Ret" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-808) A/1/3043 " downt" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3043) A/1/3974 " unt" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3974) A/1/3637 " post" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3637) A/1/3144 Start-of-word "elect", sometimes capitalized and/or with leading space Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3144) Cluster #288 A/1/671 " rest" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-671) A/1/949 " press" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-949) A/1/1138 " cross" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1138) A/1/1428 " camp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1428) A/1/2114 " trip" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2114) A/1/1064 " ent" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1064) A/1/1159 " disc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1159) A/1/1872 " refract" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1872) Cluster #395 A/0/9 This feature seems to fire for words that start with 'sc' and 'acc' like scenery, scuff, scuttle, accents, accelerate, accidents, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-9) A/1/6 " rec" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-6) A/1/534 " enc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-534) A/1/538 " inc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-538) A/1/882 "oc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-882) A/1/1347 " Ac"/" ac", sometimes in health/science Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1347) A/1/1625 " Occ" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1625) A/1/1977 " dec" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1977) A/1/2055 " prof" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2055) A/1/2795 " dev" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2795) A/1/4067 " prov"/" inv" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4067) A/1/520 "sk" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-520) Cluster #316 A/1/2872 " inst" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2872) A/1/3290 " Sc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3290) A/1/4070 " Ph" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4070) A/1/3846 " Ins" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3846) Cluster #309 A/0/194 This feature fires for words ending with "st" such as past, post, pist, wast, etc.. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-194) A/0/399 The feature fires when it sees words beginning with pl, bl, fl or tr. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-399) A/0/452 This feature seems to fire for examples of words or names that start with "Sh". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-452) A/1/36 " Sk" and few other Title Case word starts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-36) A/1/197 " cl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-197) A/1/205 " fl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-205) A/1/435 " Sh" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-435) A/1/550 " Str" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-550) A/1/660 " ch" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-660) A/1/688 " Fl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-688) A/1/708 " Cr" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-708) A/1/844 " pl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-844) A/1/859 " Wh" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-859) A/1/889 " sk" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-889) A/1/1014 " sw" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1014) A/1/1185 " Pl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1185) A/1/1187 " wh" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1187) A/1/1199 " cr" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1199) A/1/1398 " Th" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1398) A/1/1694 " sn" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1694) A/1/1974 " Sm"/"Sm" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1974) A/1/2017 " sl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2017) A/1/2091 " sh" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2091) A/1/2442 " Ch" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2442) A/1/2503 " Sw" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2503) A/1/2579 " th" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2579) A/1/2613 " bl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2613) A/1/2638 " Fr"/" fr" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2638) A/1/2898 " str" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2898) A/1/2902 " gr" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2902) A/1/3191 " st" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3191) A/1/3327 " sp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3327) A/1/3346 " Sch" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3346) A/1/3546 " Sp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3546) A/1/3598 " Br"? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3598) A/1/3806 " Cl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3806) A/1/3860 " tr" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3860) A/1/3886 " Sl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3886) A/1/4056 " gl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4056) A/1/3208 " sc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3208) A/1/2939 " ev" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2939) A/1/1239 The feature appears to fire for words containing the phonetic "arc" such as "arcades", "autoradiography", "choreocutaneous", "echocardiography", "hydrocephalus", "microautoradiography", "photocoagulation" and "ultrasonocardiography". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1239) Cluster #387 A/1/681 " vag" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-681) A/1/1825 " fib" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1825) A/1/3363 " mac" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3363) A/1/3500 " vis" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3500) A/1/1614 " pent" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1614) A/1/516 "act" as part of a word Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-516) A/1/3414 " hall" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3414) A/1/3736 Mostly " meth", some " phen" and other chemical prefixes Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3736) Cluster #159 A/1/219 This feature fires for words containing the letter combination "it" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-219) A/1/1737 " gly" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1737) A/1/2997 " glut" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2997) A/1/3432 " coll" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3432) A/1/1920 " ax" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1920) Cluster #379 A/0/91 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-91) A/1/196 " qu" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-196) A/1/826 " gu" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-826) A/1/1478 "Qu" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1478) A/1/1970 " Vo"/" vo" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1970) A/1/3169 " equ" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3169) A/1/3670 " Mult"/" mult" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3670) A/1/4016 "qu" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4016) A/1/1485 " poll" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1485) A/1/3979 " ir" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3979) A/1/3249 " semin"/" sme", predicting "ar" and forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3249) A/1/2083 " bul" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2083) A/1/2756 Start-of-word "Imag"/"fasc", predicting "ination" or forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2756) A/1/1547 " cond" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1547) A/1/1482 " spir" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1482) A/1/2155 " sem" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2155) A/1/503 "cre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-503) A/1/821 " Domin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-821) A/1/2478 " ple" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2478) A/1/1396 " pen" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1396) A/1/3418 " clim" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3418) A/1/2944 " solic" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2944) A/1/1226 " comm" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1226) A/1/860 "aph" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-860) A/1/2887 First halves of "-ulate" words Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2887) A/1/2678 " ins" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2678) A/1/921 Word beginnings " Adv"/" suff"/" coinc" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-921) A/1/2330 " att" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2330) A/1/2111 " sin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2111) A/1/858 " stabil" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-858) A/1/1810 " fin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1810) A/1/2236 " hom" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2236) A/1/3759 " throm" in medical contexts, predicting "bo-" tokens Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3759) A/1/1596 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1596) A/1/3038 " organ" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3038) A/1/305 " prima" (in "prima facie") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-305) A/1/700 "nih" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-700) A/1/3068 " syn" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3068) Cluster #401 A/0/241 The feature fires when it sees words related to photography and imaging such as photo, micrograph, videotape, photogenic, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-241) A/1/120 " mon" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-120) A/1/624 Prefixes like " aut" and " Phot", predicting "ograph" and similar suffixes Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-624) A/1/772 " immun" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-772) A/1/1025 " hyp" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1025) A/1/1743 " psych" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1743) A/1/1973 " bi" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1973) A/1/2668 " hist" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2668) A/1/3603 " cert" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3603) A/1/3934 " radi" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3934) Cluster #402 A/0/225 This feature attends to words with the beginning prefix "macro" and stem "mol" such as macromolecule, macromolecules, and macromolecular. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-225) A/1/378 " phosph" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-378) A/1/669 "ozyg" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-669) A/1/923 " gran" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-923) A/1/1363 " ur" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1363) A/1/1624 " lip" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1624) A/1/1739 " cere" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1739) A/1/1891 " lymph" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1891) A/1/3642 This feature fires when it sees medical terms ending in -itis, -emia, and -osis, as well as the word mellitus. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3642) A/1/2724 "oph" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2724) Cluster #403 A/1/445 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-445) A/1/510 " flu", in the context of bio or medical sciencey stuff, e.g. "fluorescence". Also fires on some other bio sciencey prefixes Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-510) A/1/656 Medical terms for cancers and diseases of the liver Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-656) A/1/2659 "orph" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2659) A/1/3357 "roph" mostly in "spectrophoto-" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3357) A/1/2848 "om"/"rim" largely in cell biology Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2848) A/1/3325 "olog" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3325) A/1/854 " patri" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-854) A/1/1924 " tyr"/" handic" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1924) A/1/1578 " polit" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1578) A/1/1307 " ser" prefix in seroepidemiological words Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1307) A/1/2487 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2487) Cluster #394 A/1/535 " electroly" or "analy", predicting "tic" or similar, as in "analytically" or "electrolysis" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-535) A/1/1562 " parame" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1562) A/1/3353 " plas" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3353) Cluster #400 A/0/43 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-43) A/0/408 This feature attends to words that are often found past tense or past participle verb forms. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-408) A/1/84 " altern" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-84) A/1/897 The feature attends to words ending with "ation" such as evaporation, proportional, optical or transportation. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-897) A/1/1032 Word prefixes like " incompet" and " independ" that predict endings like "ent" or "ence" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1032) A/1/1044 " expl" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1044) A/1/1107 " ped" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1107) A/1/1139 " domin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1139) A/1/1154 " illumin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1154) A/1/1774 " Contin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1774) A/1/1967 " interle" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1967) A/1/2456 " Rel" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2456) A/1/2457 "rom" in the middle of scientific/mechanical words, esp. forms of "macromolecule" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2457) A/1/2780 " lum" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2780) A/1/2928 " reass" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2928) A/1/3097 "sor" in word beginning " adsorb" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3097) A/1/3323 Mostly " quant", predicting forms of "itate" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3323) A/1/3508 " neg" / "eg" etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3508) A/1/3742 " substit" and similar word prefixes, predicting "ute" and word forms Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3742) A/1/3997 Word beginnings like " indul" and " diver" predicting suffixes like "-ge" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3997) A/1/4019 Mid-word tokens after the letter "p", esp. "lic" in "complicate"/"implicate" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4019) A/1/4058 " cran" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4058) A/1/2469 " dispar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2469) A/1/721 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-721) A/1/2415 " ann" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2415) A/1/2808 " tend" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2808) A/1/3055 " contra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3055) A/1/571 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-571) Cluster #399 A/1/60 " mot" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-60) A/1/957 Mostly " tub" with some " tum", " trib" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-957) A/1/1823 "ag" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1823) A/1/2158 "ult" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2158) A/1/2505 This feature responds to words related to light and vision, like "photoluminescence", "photoreceptor", and "photodetector". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2505) A/1/2436 Mid-word tokens after "un-" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2436) A/1/1675 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1675) A/1/316 " judic" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-316) A/1/1549 " conce" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1549) A/1/2710 Mid-word tokens after "dis" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2710) A/1/754 Usually "pe" in "Appeal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-754) A/1/3309 "Abbrev" in Abbreviations Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3309) A/1/2137 Mid-word tokens esp. after "emb" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2137) A/1/3859 " lit" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3859) A/1/3882 " Prom"/"Prom" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3882) A/1/2170 Letters after "qu" in a word Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2170) A/1/4073 " mill" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4073) Cluster #373 A/1/392 " ban" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-392) A/1/606 " par" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-606) A/1/912 " tor" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-912) A/1/1314 " gar" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1314) A/1/467 " dem" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-467) Cluster #370 A/0/92 This feature attends to words ending in an 'b' sound. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-92) A/1/164 Tokens after a single lowercase letter in a word Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-164) A/1/2162 Mid-word tokens in words starting with "s" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2162) A/1/2874 "ocal" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2874) A/1/2588 " gle" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2588) A/1/3771 This feature attends to words related to introducing, producing or inducing something. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3771) A/1/114 " ther" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-114) A/1/295 " revol" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-295) A/1/2815 " gro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2815) A/1/106 " cro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-106) A/1/2429 " cra"/" scra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2429) A/1/76 " den" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-76) A/1/1073 " che" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1073) A/1/3021 " fil" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3021) A/1/2348 " gra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2348) A/1/3688 " mo" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3688) A/1/1777 " pe" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1777) A/1/2195 " ro" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2195) Cluster #377 A/1/81 " cre" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-81) A/1/83 " ve" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-83) A/1/523 " ra" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-523) A/1/2413 " ce" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2413) Cluster #378 A/0/355 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-355) A/1/140 " lo" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-140) A/1/1188 " bo" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1188) A/1/1278 " te" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1278) A/1/2866 " le" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2866) A/1/3419 " se" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3419) A/1/3787 " ne" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3787) A/1/1993 " ma" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1993) A/1/3911 " bu" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3911) A/1/2476 " une" or " sa"? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2476) A/1/2271 " Los"/"Los" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2271) Cluster #182 A/1/155 " su" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-155) A/1/1358 " Je" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1358) A/1/3883 " du" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3883) A/1/479 " de" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-479) A/1/1477 " av" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1477) Cluster #193 A/0/76 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-76) A/0/206 "la" in french and spanish Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-206) A/1/597 Italian or Spanish: Feminine articles " una"/" la" in Spanish and maybe Italian / other languages? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-597) A/1/1251 Spanish: " de" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1251) A/1/1965 Spanish: Word beginnings? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1965) A/1/379 " di" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-379) Cluster #208 A/0/61 This feature seems to fire when it sees text in Portuguese or Brazilian Portuguese. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-61) A/0/487 The 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-487) A/1/30 Spanish: " que", usually Spanish? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-30) A/1/434 " no" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-434) A/1/542 Spanish: " el" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-542) A/1/725 Spanish: word endings Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-725) A/1/944 Spanish or Portuguese: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-944) A/1/1337 Spanish: " en" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1337) A/1/1997 Spanish: Punctuation, esp. colons Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1997) A/1/2109 Portuguese tokens, esp. " prepar", predicting "ão" and endings Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2109) A/1/2244 Spanish: " los" and other articles? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2244) A/1/2331 Spanish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2331) A/1/2502 Portuguese: Prepositions/articles Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2502) A/1/2989 Portuguese: Strong tokens with accented and distinguishing Portuguese characters Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2989) A/1/3024 Portuguese: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3024) A/1/3140 Spanish or Portuguese: " a" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3140) A/1/3656 Spanish: "reg" and other mid-word tokens? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3656) A/1/3716 Spanish: "Yo" / "adie" / ... ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3716) A/1/3836 Short word fragments esp. just one capital letter Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3836) A/1/833 Spanish: Space+letter tokens? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-833) Cluster #199 A/0/14 This feature appears to fire when French is present. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-14) A/0/44 This feature fires when it sees accented characters, like é, ã, ö, â, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-44) A/0/182 This feature seems to fire on apostrophes in French text. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-182) A/0/255 This feature fires when it sees Spanish words related to government, legal, and business topics such as politics, legislation, petitions, solicitation, requests, and registration. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-255) A/1/150 " est"/" Est" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-150) A/1/436 French: " les"/" des" (the/some) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-436) A/1/544 French: Apostrophe as contraction Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-544) A/1/956 Spanish or Portuguese: Some -du word beginnings? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-956) A/1/981 French: é in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-981) A/1/991 French: " qui" but only in the that sense, not who Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-991) A/1/1050 French: predicts é|e|è Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1050) A/1/1168 French: " é" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1168) A/1/1288 French: "è" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1288) A/1/1892 French: Prepositions? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1892) A/1/2016 French: " :", and other punctuation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2016) A/1/2042 French: " dé"/" ré" prefix only Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2042) A/1/2146 French: Names / proper nouns, perhaps in a bibliography? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2146) A/1/3002 French: Prepositions about containing/linking Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3002) A/1/3008 French: Predicts prepositions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3008) A/1/3207 "\u00e2" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3207) A/1/3596 French: " à" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3596) Cluster #209 A/0/184 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-184) A/1/2034 Italian: Formal governmental (Not sure why it it's matching on a subset of the Italian words/characters though) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2034) A/1/2156 Spanish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2156) A/1/2250 Italian: elements of verb conjugation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2250) Cluster #200 A/0/455 This feature attends to text in Romanian. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-455) A/1/9 Romanian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-9) A/1/632 Romanian: " \u00ee" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-632) A/1/846 Accented, esp. acute, characters Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-846) A/1/3923 Latin: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3923) A/1/734 Indonesian/Malaysian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-734) Cluster #211 A/0/159 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-159) A/1/1657 Turkish, Basque, Hawaiian, Maori, and other Polynesian languages: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1657) A/1/1912 Icelandic: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1912) A/1/623 Vietnamese & Romance Languages: Text near UTF-8 as latin1 mojibake Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-623) Cluster #212 A/0/36 This feature is attending to Slavic languages, mainly Polish and Czech. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-36) A/0/294 This feature attends to Hungarian words and some other Hungarian-like examples from nearby languages. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-294) A/1/1141 "\u00e1" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1141) A/1/1200 "," in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1200) A/1/1627 Turkish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1627) A/1/1772 Estonian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1772) A/1/2201 Lithuanian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2201) A/1/2360 Slovenian: Text or names? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2360) A/1/2651 Slovenian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2651) A/1/2676 Hungarian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2676) A/1/2792 Slovenian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2792) A/1/2801 Slavic: "," in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2801) A/1/3348 Polish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3348) A/1/3361 Latvian: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3361) A/1/3537 Vietnamese: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3537) A/1/3648 Czech: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3648) A/0/180 This feature attends to punctuation such as periods, quotation marks, newlines. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-180) A/1/3329 Finnish: Word beginnings? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3329) A/0/162 This feature fires when it sees German words that start with an "s" followed by a consonant. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-162) Cluster #205 A/0/339 This feature seems to be firing for Finnish words and sentences. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-339) A/1/1052 Finnish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1052) A/1/2570 Finnish and German?: Consonants after umlauts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2570) A/1/2897 Finnish: Umlauts (and some surrounding text?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2897) A/1/3413 Finnish: tokens esp. with "-ist" or "-ll" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3413) A/1/2153 Swedish?: Letters or tokens with diacritics Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2153) A/1/1112 [Ultralow density cluster] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1112) A/1/917 Dutch: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-917) Cluster #213 A/0/469 This 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. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-469) A/1/321 Swedish: Punctuation Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-321) A/1/1373 Danish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1373) A/1/2464 Swedish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2464) A/1/3046 Danish: Politics Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3046) Cluster #196 A/0/493 This feature seems to activate for German text, particularly German grammar such as noun declensions. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-493) A/1/177 German: " der", plus other articles/prepositions weakly Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-177) A/1/334 German: Text tokens ending in "e" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-334) A/1/358 German: Consonants at start of word Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-358) A/1/838 German: Prepositions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-838) A/1/1695 " und" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1695) A/1/2410 German:? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2410) A/1/2601 German: Commas Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2601) A/1/2637 German: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2637) A/1/2652 German: Pronouns ("ihre", "die", "dieser") Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2652) A/1/2661 German: Modal verbs Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2661) Cluster #214 A/1/1951 Dutch: " wa"/" na"? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1951) A/1/2065 Dutch: Some word beginnings? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2065) A/1/2899 Danish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2899) A/1/2032 Swedish: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2032) A/1/103 LaTeX length names, particularly "side" in "oddsidemargin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-103) A/1/2576 "ensuremath" in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2576) Cluster #116 A/0/35 This feature fires for math expressions, in particular fractions displayed using LaTeX. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-35) A/0/82 This feature attends to mathematical symbols. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-82) A/1/267 LaTeX mathbb font Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-267) A/1/427 LaTeX operators with a number over another (frac, binom, dfrac) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-427) A/1/914 LaTeX macro names esp. accepting one short argument Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-914) A/1/1136 LaTeX environment "begin" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1136) A/1/1889 LaTeX text modifiers in equations (inc. mathfrak, sqrt, overline, boldsymbol) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1889) A/1/2226 This feature seems to fire on LaTeX formatting and math markup. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2226) A/1/2590 This feature fires on TeX/LaTeX formatting such as equations, labels, sections, tables, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2590) Cluster #3 A/0/175 This feature attends to mathematical symbols and expressions. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-175) A/0/368 This feature fires when it encounters LaTeX commands and math symbols. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-368) A/1/112 Code: "\" in paths or string escapes Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-112) A/1/193 LaTeX: Inline modifiers Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-193) A/1/222 LaTeX: Predicting an operator? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-222) A/1/262 Code: "\" in RTF source code? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-262) A/1/363 LaTeX: Predicting command with alternate tokens Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-363) A/1/848 LaTeX: Punctuation clusters ending in \, predicting a command Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-848) A/1/1123 LaTeX: " $\", predicting greek letters in formulae Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1123) A/1/1277 LaTeX: Predicting command Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1277) A/1/1506 LaTeX: Bracket and "\", predicting command (often formatting) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1506) A/1/1911 LaTeX: "(\" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1911) A/1/1917 LaTeX: Superscripted or subscripted backslash, predicting macros in superscripts/subscripts' Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1917) A/1/2313 LaTeX: Predicting escaped characters Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2313) A/1/2737 LaTeX: "\" and " \" - left/right delimiters, etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2737) A/1/3205 LaTeX: " $$\\\" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3205) A/1/3471 LaTeX: "^{\" and "_{\", introducing superscripted/subscripted commands Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3471) A/1/3517 Back Slash in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3517) A/1/3853 LaTeX: Predicting equation symbols Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3853) A/1/3891 LaTeX: "{\" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3891) A/1/3894 LaTeX: Subscript with escape Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3894) Cluster #109 A/0/173 This feature fires for LaTeX markup, specifically for commands that start an environment or matrix. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-173) A/1/529 LaTeX font setting, primarily \rm Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-529) A/1/3789 "usepackage" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3789) Cluster #77 A/0/158 The feature attends to mathematical notation and formulas. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-158) A/0/205 This feature fires on mathematical notation. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-205) A/1/108 LaTeX contents of curly-braced superscripts or subscripts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-108) A/1/356 LaTeX measurements (probably units) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-356) A/1/455 "aligned" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-455) A/1/2452 Contents, esp. digits, of curly braces in superscripts or subscripts, predicting close braces Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2452) A/1/3340 LaTeX letters in \mathcal and a few other fonts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3340) A/1/3552 Numbers after open curly braces in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3552) A/1/1798 Code: Letters in backslash escapes (n in \n etc.) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1798) A/1/2927 LaTeX fraction numerators Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2927) A/1/3384 "69" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3384) A/1/723 LaTeX: macro names for spacing like "qquad" or dots like "ldots" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-723) Cluster #88 A/0/93 This feature attends to text written in LaTeX. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-93) A/1/633 "}" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-633) A/1/840 " \\\\\\\"/"\\\\\\\" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-840) A/1/1186 Hexadecimal digits in \x escapes Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1186) Cluster #98 A/0/185 This feature appears to attend to mathematical equations with numbers, particularly those with the value of 1 or 2. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-185) A/1/788 Numbers after "=" or similar operators, e.g. \approx Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-788) A/1/2401 LaTeX exponents (after "^"), esp. 2 Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2401) A/1/2501 Numbers after "_" (subscripts), often predicting superscripts Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2501) Cluster #68 A/1/1285 Physics notation, esp. near square brackets (hints of relativity) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1285) A/1/2221 LaTeX: numbers in matrices/tables/etc. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2221) A/1/2785 This feature attends to mathematical notation with formulas and equations. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2785) Cluster #117 A/0/111 This feature fires when it sees mathematical notation, specifically symbols used in calculus and algebra. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-111) A/0/161 This feature attends to mathematical formulas written in LaTeX. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-161) A/0/499 This feature fires for mathematical variables. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-499) A/1/137 LaTeX big delimiters Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-137) A/1/189 LaTeX: "right" command Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-189) A/1/1035 Greek letters in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1035) A/1/1576 Greek letter macros in LaTeX math mode Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1576) A/1/1829 LaTeX: short, mostly single-letter variables Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1829) A/1/1929 This feature fires when it sees mathematical equations and LaTeX-style notation. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1929) A/1/2056 LaTeX trigonometric functions Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2056) A/1/2512 LaTeX left delimiter opening Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2512) A/1/2750 \to in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2750) A/1/2800 LaTeX operators (sum, prod, int, etc.) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2800) A/1/2920 LaTeX Math (general) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2920) A/1/3057 Feature for predicting "_{" and similar in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3057) A/1/3333 Single-letter tokens/variables in LaTeX math mode Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3333) A/1/3574 Greek letters in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3574) A/1/3752 One-letter variables, often subscripts, in LaTeX math (predicting superscripts) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3752) A/1/3940 Space + capital letter in LaTeX often a variable after a macro like "\in" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3940) A/1/4044 "partial" in the LaTeX macro "\partial" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4044) A/1/697 Numbers in parentheses, often comma-separated tuples, in LaTeX math mode Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-697) A/1/2295 Single-digit numbers in LaTeX math Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2295) A/0/144 This feature attends to mathematical formulas. It activates on variables, mathematical constants and operators, as well as formatting and punctuation commonly used in equations. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-144) A/1/2864 This feature fires in response to mathematical formulae, particularly those containing mathematical symbols such as Greek letters. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2864) A/1/596 Latin-alphabet identifiers in LaTeX Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-596) A/1/3258 LaTeX: variables in parenthesized expressions or pairs in inline math mode Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3258) A/0/167 This feature appears to fire for code examples and programming language keywords, especially in C/C++. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-167) Cluster #75 A/0/275 This feature attends to capital letter A at the beginning of sentences. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-275) A/0/325 This feature fires when it sees a single capital letter followed by a period, most often at the beginning of a sentence. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-325) A/1/654 "S" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-654) A/1/806 This feature attends to acronyms, initialisms and abbreviations in the text. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-806) A/1/1207 'g' in 'e.g.', predicts the next period Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1207) A/1/1773 " A" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1773) A/1/1998 " A" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1998) A/1/2243 This feature attends to US Court case citations. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2243) A/1/3625 "A" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3625) A/1/4014 "L" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4014) Cluster #121 A/1/784 " X" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-784) A/1/1429 " N" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1429) A/1/2125 " D" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2125) A/1/2776 " V" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2776) Cluster #70 A/0/213 This feature attends to words that contain the letter "b". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-213) A/1/300 " d" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-300) A/1/438 " e" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-438) A/1/476 " w" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-476) A/1/530 " k" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-530) A/1/738 " m" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-738) A/1/825 " h" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-825) A/1/910 " p" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-910) A/1/1079 " r" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1079) A/1/1779 " j" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1779) A/1/1790 "f" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1790) A/1/2108 " b" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2108) A/1/2265 " t" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2265) A/1/2332 " s" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2332) A/1/2335 " l" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2335) A/1/2667 " g" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2667) A/1/3371 " c" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3371) A/1/3579 " v" usually beginning a word Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3579) A/1/3730 " n" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3730) A/1/3993 " o" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3993) Cluster #120 A/0/26 This feature fires when it sees initials, especially two initials with a period between them. It attends to proper nouns and names in abbreviation format. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-26) A/0/137 This feature fires on the letter D in text. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-137) A/0/296 This feature fires in response to proper nouns that start with "F" or some other capital letters. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-296) A/0/395 This feature is firing for instances of "X.P" and "P.X" where "X" represents uppercase Roman numerals and "P" represents single digit page numbers. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-395) A/0/479 The feature fires when it sees capital M followed by a capital letter. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-479) A/1/11 " M" esp. shortly after <EOT> Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-11) A/1/323 " Z" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-323) A/1/1057 " P" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1057) A/1/1871 " T" in titles/names/English Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1871) A/1/3052 " C" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3052) A/1/3107 " Y" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3107) A/1/3159 " J" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3159) A/1/3289 " G" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3289) A/1/3406 " W" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3406) A/1/3408 " F" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3408) A/1/3428 " L" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3428) A/1/3939 " R" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3939) A/1/732 " H" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-732) A/1/1740 " K" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1740) A/1/1786 " B" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1786) A/1/748 This feature fires for character names that start with a soft C, D, F, J, K, M, N, P, R, T, V, W, Z sound. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-748) A/1/1010 " S" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1010) A/1/1718 "C" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1718) A/0/350 This feature attends to capital S followed by 4-9 capital letters. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-350) Cluster #93 A/0/278 This feature fires for names and words starting with the letter E. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-278) A/1/1377 " O" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1377) A/1/3351 " U" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3351) A/1/4008 " E" in [context] Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-4008) A/1/3528 "B" esp. after "-" or "/" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3528) Cluster #1 A/0/397 This feature attends to Chinese sentences and phrases. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-397) A/1/374 Chinese: Technical Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-374) A/1/900 Chinese: Simplified Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-900) A/1/2000 Chinese: Obscure characters / dictionaries / transliterated names Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2000) A/1/2030 Japanese: Numbers (dates and laws?) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2030) Cluster #2 A/0/304 This feature fires on Japanese text. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-304) A/1/698 Japanese: Period, end-of-sentence character Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-698) A/1/1260 Chinese: Negations / "but" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1260) A/1/2730 Japanese: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2730) Cluster #8 A/1/1671 Japanese and Chinese: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-1671) A/1/2239 Japanese: "そ"? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2239) A/1/3655 Japanese, Chinese, and Korean: ? Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3655) Cluster #5 A/1/2602 Japanese: 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) Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2602) A/1/2986 Japanese: The "です desu" copula that means "is". It represents "X is Y" in "X Y desu" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2986) A/1/3112 Japanese: Suffix that noun-ize a verb, similar to "ing" suffix in english Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3112) A/1/3605 Japanese: A form of the "to do" verb prefix. Usually passive tense, occasionally "can/able to do", and sometimes "is doing" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3605) A/1/3927 Japanese: Hiragana "e" character, predicting things likely to come after that Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3927) A/1/3965 Japanese: "な" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3965) Cluster #9 A/0/343 This feature fires when it sees Japanese text. Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-343) A/1/63 Japanese: Hiragana particle that follows a noun, representing it was a direct or indirect object Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-63) A/1/380 Japanese: 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 Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-380) A/1/421 Japanese: Specialized - game/IRC chats Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-421) A/1/964 Japanese: Katakana characters representing foreign-loan words Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-964) A/1/2850 Japanese: Hiragana ha/wa character, that is particle representing the phrase before it is a "topic" Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-2850) A/1/3829 Japanese: Hiragana "no" that represents possessive conjunction, similar to "apostrophe-s" in English Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a1.html#feature-3829) Cluster #22 A/0/267 This 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". Zoom to [↗ View details](https://transformer-circuits.pub/2023/monosemantic-features/vis/a0.html#feature-267) A/0/331 This feature attends to the use of the preposition to'' in a variety of functional contexts.
Zoom to ↗ View details
A/1/297
" to" in [context]
Zoom to ↗ View details
A/1/390
" to" in [context]
Zoom to ↗ View details
A/1/975
" to" in [context]
Zoom to ↗ View details
A/1/990
" to" in [context]
Zoom to ↗ View details
A/1/1541
" to" in [context]
Zoom to ↗ View details
A/1/2169
" To"
Zoom to ↗ View details
A/1/2439
" to" in [context]
Zoom to ↗ View details
A/1/2543
" to" in [context]
Zoom to ↗ View details
A/1/3613
" to" in [context]
Zoom to ↗ View details
A/1/3854
" to" in [context]
Zoom to ↗ View details
A/1/3484
"to" in [context]
Zoom to ↗ View details
A/1/2240
" to" in [context]
Zoom to ↗ View details
Cluster #23
A/0/226
The feature fires when the word "to" is used to indicate an infinitive verb.
Zoom to ↗ View details
A/1/1262
" to" in [context]
Zoom to ↗ View details
A/1/1364
" to" in [context]
Zoom to ↗ View details
A/1/1802
" to" in [context]
Zoom to ↗ View details
A/1/1832
" to" in [context]
Zoom to ↗ View details
A/1/2063
" to" in [context]
Zoom to ↗ View details
A/1/2191
" to" in legal
Zoom to ↗ View details
A/1/2361
" to" in [context]
Zoom to ↗ View details
A/1/2653
[Ultralow density cluster]
Zoom to ↗ View details
A/1/2666
" to" in [context]
Zoom to ↗ View details
A/1/2688
" to" in [context]
Zoom to ↗ View details
A/1/2960
" to" in [context]
Zoom to ↗ View details
A/1/3018
" to" in [context]
Zoom to ↗ View details
A/1/3297
" to" in [context]
Zoom to ↗ View details
A/1/3778
" to" in [context]
Zoom to ↗ View details
A/1/3805
" to" in infinitive phrases (verb " to" verb)
Zoom to ↗ View details
A/1/4011
" to" in [context]
Zoom to ↗ View details
A/1/2172
" to" in [context]
Zoom to ↗ View details
A/1/3956
" to" in [context]
Zoom to ↗ View details
A/1/1047
" to" in [context]
Zoom to ↗ View details
A/1/905
" to" in [context]
Zoom to ↗ View details
scroll zoom homezoom inzoom out
A/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/1/4082A 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.
In 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.
In 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":
Crucially, 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.
| | Transformer | Sparse Autoencoder |
| --- | --- | --- |
| Layers | 1 Attention Block
1 MLP Block (ReLU) | 1 ReLU (up)
1 Linear (down) |
| MLP Size | 512 | 512 (1×) – 131,072 (256×) |
| Dataset | The Pile
(100 billion tokens) | Transformer MLP Activations
(8 billion samples) |
| Loss | Autoregressive Log-Likelihood | L2 reconstruction
There 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 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. If 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. Instead, we decompose the activation vector \mathbf{x}^j as a combination of more general features which can be any direction: \mathbf{x}^j \approx \mathbf{b} + \sum_i f_i(\mathbf{x}^j) \mathbf{d}i (1) where \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. In our sparse autoencoder setup, the feature activations are the output of the encoder f_i(x) = \text{ReLU}( W_e (\mathbf{x} - \mathbf{b}_d) + \mathbf{b}_e )_i, where 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 for full notation.) If 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 on feature universality suggest that features have some existence beyond individual models.
To 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 .
A 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.
Suppose 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:
A feature decomposition satisfying these criteria would allow us to:
Of 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.
In Toy Models of Superposition , 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. Initially, 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. Unfortunately, having spent a significant amount of time investigating this approach, we have ultimately concluded that it is more fundamentally non-viable. In 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. To 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. We can calculate the cross-entropy loss achieved by this model in a few cases:
Because 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. This 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. Based 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. Models 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. Note, 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.")
There 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. For this reason, we seek a decomposition which is sparse and overcomplete. This is essentially the problem of 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.) It'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, 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. Despite 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", 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.
We briefly overview the architecture and training of our sparse autoencoder here, and provide further details in Basic Autoencoder Training. 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. We 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? loss plus an L1 penalty to encourage sparsity. In 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). Second, 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, 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. For readers looking to apply this approach, we supply an appendix with Advice for Training Sparse Autoencoders.
Usually 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. We began by looking for an information-based metric, 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.) Thus we ended up using a combination of several additional metrics to guide our investigations:
Interpreting 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. 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. We think it would be very helpful if we could identify better metrics for dictionary learning solutions from sparse autoencoders trained on transformers.
We 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:
We 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. We 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, as we can e.g. compare features learned from the “A” and “B” transformers and see how similar they are.
Throughout 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. 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. Sometimes, 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".
We 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. We suggest beginning with the interface for A/1, which we discuss the most.
These 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:
Our interface also allows users to search through features:
Additionally, we provide a second interface displaying all features active on a given dataset example. This is available for a set of example texts.
The 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. The features we study respond to
For each learned feature, we attempt to establish the following claims:
To 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.
In 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 and Feature Splitting.
We 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 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.
We 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, would be difficult, and we analyze them in other ways in the following section.
The first feature we'll consider is an Arabic Script feature, A/1/3450. 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.
Our 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.
We 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) 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.
The 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:
Regardless, 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.
In 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, fires. There are several additional features that fire on Arabic and related scripts (e.g. A/1/1466, A/1/3134, A/1/3399) 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 activates on the first token comprising the character. To see how these features collaborate, we provide an alternative visualization showing all the features active on a snippet of Arabic text. We consider such interactions more in the Phenomenology section below. Nevertheless, 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.
Because 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.
We 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.
Each 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).
This 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:
. (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.
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 to its maximum observed value and see how that changes the samples:
This 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 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.)
We 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.
It 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) 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.
Logit 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.
Finally, scatter plots and correlations suggest the similarities between A/1/3450 and the neuron are non-zero, but quite minimal.By analogy, this also applies to B/1/1334. 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.
We 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.
We 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.
We search for a similar feature in B/1, 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. and find B/1/1334 (corr=0.91), which is strikingly similar:
This 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:
The effects of ablating this feature are also consistent with this (see the visualization for B/1/1334).
To more systematically analyze the similarities between A and B, we look at scatter plots comparing the activations or logit weights:
The activations are strongly correlated (Pearson correlation of 0.91), especially in the main Arabic mode.
The 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.
We now consider a DNA feature, A/1/2937. 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 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.
We 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.
While 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.
We 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 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, it is the only feature active.
The 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.
The most similar neuron to A/1/2937, as measured by activation correlation, is A/neurons/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, also has no DNA sequences in its top activating examples.
A/1/2937 has a correlated feature (corr=0.92) in run B/1, B/1/3680. 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.
We now consider a base64 feature, A/1/2357. 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.
This is not the only feature active in base64 contexts, and in a section below 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.
Turning 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.)
There 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.
The Pearson correlation between the computational proxy and the activity of A/1/2357 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.
A/1/2357 has a correlated feature (corr=0.85) in run B/1, B/1/2165. It also has high activation specificity for base64 strings:
Like A/1/2357, B/1/2165's logit weights have a second mode corresponding to base64 tokens:
Correlations and scatter plots are also consistent with them being very similar features:
Note 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.
Looking at the neuron in model A that most correlates with this feature: A/neurons/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.:
The logit weights suggest it somewhat increases base64 tokens, but is much more focused on upweighting other tokens, e.g. filename endings.
The activation and logit correlations are consistent with this neuron helping represent the same feature, but largely doing other things.
Another interesting example is the Hebrew feature A/1/416. Like the Arabic feature, it's easy to computationally identify Hebrew text based on Unicode blocks.
A/1/416 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).
Its 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.
The Pearson correlation of the Hebrew script proxy with A/1/416 is 0.55. Some of the failure of sensitivity may be due to a complementary feature A/1/1016 that fires on
\xd7 and predicts the bytes that complete Hebrew characters' codepoints.
There doesn't appear to be a similar neuron. The most correlated neuron in model A is A/neurons/489 (corr=0.1), which has low activation and logit specificity. Consider the following activation and logit correlation plots:
To cross-validate this, we also searched for any neuron where the main Hebrew Unicode block appeared in the top dataset examples. We found none.
A/1/416 has a correlated feature in the B/1 run, B/1/1901 (corr=0.92) that has significant activation specificity:
Logit weights have a second mode, as before:
Activation and logit weight correlations are again consistent:
If 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. There 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. We 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. We exclude both these groups of features from further analyses.
In 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.
At 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 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.
In 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.
Unfortunately, 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.
We 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.
To 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.
Like 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 for additional information including the use of importance scoring to precisely counter-weight the bias of providing tokens the feature fires for.
In 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.
In 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.
In 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.
Higher-activating feature intervals were more consistent with our interpretations than lower-activating ones. In particular:
It 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.
Our manual and automated interpretability experiments have a few caveats:
Based 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 and the corresponding neurons. 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.
We 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?
One way to partly get at this question is to ask how much of the loss is explained by our features. For A/1, 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/5 (n_learned_sparse=131,072, l1_coefficient=0.004) recovers 94.5% of log-likelihood loss.
These 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 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.
In 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.
Overall, 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.
A 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. To 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, 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 or code). 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. To assess the interpretability of the downstream feature effects, we again use the three main approaches of the previous section:
The 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.
Ultimately, 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, and the 2019 ICML workshop on phenomena in deep learning.) One 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 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?" We 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.
What kinds of features do we find in our model?
One 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, < in HTML – A/0/20).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, 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.
Another 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). 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.
Finally, 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 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.
One 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.
To 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.
We 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.
We 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.
In 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 and collapsing of correlated features), 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.
If 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), or else highly-specific "split" features which do in fact make sense if analyzed closely (see "Bug" 2: Multiple Features for a Single Context). 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.
In 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.
What 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.
It'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) 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) has much more specific topical predictions (e.g. the dataset, the classifier). Likewise, our abstract algebra and topology feature (A/2/4878) supports the quotient and the subgroup, and the gravitation and field theory feature (A/2/2609) supports the gauge, the Lagrangian, and the spacetime.
"Bug" 1: Single-Token Features
When 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 .
We 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.
"Bug" 2: Multiple Features for a Single Context
We 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 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, 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?
In A/0 (with 512 features), the story is simple. There is only one base64-related feature, A/0/45, which seems to activate on all tokens of base64-encoded strings. But in A/1, that feature splits into three different features whose activations seem to jointly cover those of A/0/45:
Two of these features seem relatively straightforward. A/1/2357 seems to fire preferentially on letters in base64, while A/1/2364 seems to fire preferentially on digits.
Comparing 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.
We 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.
But what about the third feature, A/1/1544? 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 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 contain substrings which decode as ASCII, while none of the top activating examples for A/1/2357 or A/1/2364 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 of ASCII characters. Thus, we select substrings which – when decoded with the python base64 library – contain the maximal number of printable ASCII characters.
This 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.
It'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.
One of the biggest "meta questions" about features is whether they're universal – 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. Earlier, 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. We 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.
To compare features from different models, we need model-independent ways to represent a feature.
One 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.
A 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.
These two notions of similarity correspond to the correlations of the points in the two scatter plots we used when analyzing individual features earlier. We've reproduced the plots for the Arabic feature below:
For each feature in run A/1, we find the closest feature by activation similarity in run B/1, 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 having a median activation correlation of 0.72 with the most similar feature from B/1. (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.
A 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?
Some 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.
In the scatterplot below, we find that this kind of disagreement is widespread.
The most dramatic example of this disparity is for the features A/1/3949 and B/1/3321, 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.
Zooming 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.
We 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.
In 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.
So 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. The 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), hexadecimal (A/0/119), and all caps (A/0/317) features here. Discussion of some of these neurons can be found in Section 6.3.1 of the SoLU paper. We 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 of this phenomenon), we find a similar German detector (e.g. A/0/493) and similar title case detectors (e.g. A/0/508). Likewise, we find a number of features similar to Gurnee et al. , including a "prime factors" feature (A/4/22414) and a French feature (A/0/14). At 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), Canada feature (A/3/13683), Africa feature (A/3/14490), and Israel-Palestine feature (A/3/739) 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 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.
One 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.
The 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.
It's worth noting that these examples are from A/0, 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.
Let'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) activates on the all caps text tokens, the other (A/0/358) on underscores:
This 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.)
For example, Tamil Unicode characters (block U+0B80–U+0BFF) are typically split into two tokens. For example, the character "ண" (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.
A 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.
Let's now consider a very simple four node system which models HTML. The "main path" through it is:
A prototypical sample this might generate is something like <div>\n\t\t<span>.
The full system can be seen below:
Keep in mind that we're focusing on the A/0 features where this is very simple – if we looked at A/1, we'd find something much more complex! One particularly striking shortcoming of the A/0 features is that they don't describe what happens when A/0/0 emits a token like
href, which leads to a more complex state.
It'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:
A prototypical sample this might generate is something like
<nickonia_> lol ubuntu ;). Presumably the Pile dataset heavily represents IRC transcripts about linux.
One 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). 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.
This 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.
Superposition 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 of Toy Models of Superposition where we discuss these topics in depth, and also to our essay Distributed Representations: Composition & Superposition . 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.
Since we published Toy Models of Superposition there has been significant further work attempting to better understand superposition. We briefly summarize below. Is 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. When 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. Memorization – 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 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.
There 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. This 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. Framed 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. Similarly, 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. These 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).
Our work builds on a longer tradition of using dictionary learning and sparse autoencoders to decompose neural network activations. Early work in this space focused on word embeddings and other non-transformer neural networks. Faruqui et al. and Arora et al. 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. More 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. At 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. In 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). 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.
Coming 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.
This 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.
Moreover, 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).
These 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 observed above.)
Nevertheless, 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.
Finally, 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.
One 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"). But 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! To 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". The more intuitive way to represent this would instead be a "compositional code" – representing "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:
If 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. However, 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.
Scaling 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. Scaling 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. How 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 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. Scalability 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). Algorithmic 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 . Attentional 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 and July Updates). If this is true, addressing this may become a future bottleneck for the mechanistic interpretability agenda. Theory 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.
Inspired by the original Circuits Thread and Distill's Discussion Article experiment, 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.
Neel Nanda is an external mechanistic interpretability researcher. This is a summary of a blog post replicating and extending this paper, with an accompanying tutorial to load some trained autoencoders and practice interpreting a feature.
The 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.
I've open sourced two trained autoencoders and a tutorial for how to use them, and how to interpret a feature. I've also open sourced a (very!) rough training codebase, along with some of the implementation details that came up, and tips for training your own.
I investigated how sparse the decoder weights are in the neuron basis 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.
I also exhibit some case studies of the features I found, like a title case feature, and an "and I" feature
I didn’t find any dead features, but more than half of the features form an ultra-low frequency cluster (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!
One 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, 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.
General 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. Sparse 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. Analysis 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. Interface – 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.
Feature 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. Manual 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. Automated 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. Feature Splitting – Josh Batson, Adam Jermyn, and Chris Olah analyzed feature splitting. Shan Carter produced the UMAPs of features. Universality – Josh Batson analyzed universality, with engineering support from Trenton Bricken and Tom Henighan. "Finite State Automata" – Chris Olah analyzed "Finite State Automata".
Writing – 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. Illustration – Diagrams were created by Chris Olah, Josh Batson, Adam Jermyn, and Shan Carter, based on data generated by themselves and others.
Early 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. Sparse 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.
Support – 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. Leadership – 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.
We 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. We 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. Our 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.
Please cite as:
Bricken, et al., "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning", Transformer Circuits Thread, 2023.
BibTeX Citation:
@article{bricken2023monosemanticity,
title={Towards Monosemanticity: Decomposing Language Models With Dictionary Learning},
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},
year={2023},
journal={Transformer Circuits Thread},
note={https://transformer-circuits.pub/2023/monosemantic-features/index.html}
}
In 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. Since 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. But how can we estimate the different terms? Let's consider each part separately.
To compute P(s), we use the unigram distribution over tokens and compute P(s) = \prod_{t\in s} p(t).
For 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).
We model DNA as a random string of [ATCG], each with probability 1/4.
(Although we don't use them in this draft, we also found this approach helpful for hexadecimal, binary, and similar features.)
We also construct proxies for languages written in scripts that use distinctive unicode blocks. We'll use Arabic as an example of this. For 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, U+0750–U+077F). 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.
We 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.
Our rubric for scoring how interpretable features are has the following instructions:
The 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.
While 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.
We 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.
As 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.
Ideally, 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.
Many of the feature density histograms are bimodal. For example, consider the feature density histogram for A/16:
This 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 .
In 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.
Even 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.
The 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.
Below, 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.
The 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. We 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.
To 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. For 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.
Our 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 . Formally, 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: \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} Note 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. Our 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. Here is a summary of the modifications sorted both between and within sections in rough order of importance
Testing the autoencoder on toy problems with a bias, we found that it would fail to produce the “bounce plots” we described previously 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 .
It 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. For example, in "Bug" 2: Multiple Features for a Single Context 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. One 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.
Over 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) 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. An 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. The 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. . Better resampling strategies is an active research area of ours. The approach used in this work is as follows:
This 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. This 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.
We 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). We also tried annealing our learning rate over the course of training but found that this did not further increase performance.
Recall 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.
The 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. We most frequently look at the following proxy metrics:
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.
Feature Density Histograms: Specific metrics from these histograms include:
The number of alive features outside of the ultralow density cluster
The minimum feature density at which we see a significant number of non-ultralow-density-cluster features.
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.
\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.
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.
We 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.
The L^0 norm and Reconstructed Transformer NLL metrics, taken together, display a human-interpretable tradeoff between sparsity and explained behavior:
Each 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.
Visualizations 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.
We 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.
Here we provide more details on the implementation of our automated interpretability before explaining importance scoring and providing additional results.
In 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.
Using 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.
Note 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.
For 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.
We 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.
As 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.
It 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.
Features 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).
Meanwhile, 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.
Note 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.
The 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.
Importance 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.
The 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.
The 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).
Claude 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.
Here 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.
To 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, A/0, randomized model features, neurons.
[ P][attern].[↩]P features by manual inspection rather than by cosine similarity.[↩]ICAgICAg which this feature particularly responds to, and corresponds to six spaces in a row.[↩]