{ "source_file": "0930", "source_platform": "知乎专栏", "author": "刘聪NLP", "title_listed": "大模型微调经验总结与项目更新", "keywords": "LLM,微调,ChatGLM,SFT", "domain": "大模型算法", "content_type": "工程实践", "is_paid_flag": "否", "url": "https://zhuanlan.zhihu.com/p/620885226", "publish_date": "2023", "corpus_stage": "后训练", "notes": "ChatGLM 微调项目重构，支持 Freeze/LoRA/P-Tuning/全参", "id": "e0eda9767a9f", "collected_at": "2026-10-02T18:38:50", "title": "大模型LLM-微调经验分享&总结 - 知乎", "body_markdown": "[![](https://zhuanlan.zhihu.com/p/620885226)](javascript:void\\(0\\))\n![ZhiHu logo](../images/body-e1fa91dbdb48b8055184.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n大模型LLM-微调经验分享&总结\n首发于[AIGC](https://www.zhihu.com/column/c_1625207909599617024)\n切换模式\n![大模型LLM-微调经验分享&总结](../images/000.img)\n# 大模型LLM-微调经验分享&总结\n[![刘聪NLP](../images/024.jpg)](https://www.zhihu.com/people/LiuCongNLP)\n[刘聪NLP](https://www.zhihu.com/people/LiuCongNLP)\n[​![](../images/body-2dd3e5c360100c555650.png)](https://www.zhihu.com/question/510340037)\n新知答主\n[收录于 · 人工智能研究](https://www.zhihu.com/column/c_1332763253097734144)\n1731 人赞同了该文章\n## **写在前面**\n大型语言模型横行，之前非常焦虑，现在全面拥抱。目前也有很多开源项目进行大模型微调等，笔者也做了一阵子大模型了，特此来介绍一下ChatGLM-6B模型微调经验，并汇总了一下目前开源项目&数据。笔者与很多人微调结论不同，本人在采用单指令上进行模型微调，发现模型微调之后，**「并没有出现[灾难性遗忘](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E7%81%BE%E9%9A%BE%E6%80%A7%E9%81%97%E5%BF%98&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLngb7pmr7mgKfpgZflv5giLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.m2m3IzZUhVIJqj3GlQeWAfn-TCtYst79-om2wR_2qgk&zhida_source=entity)现象」**。\n[](https://link.zhihu.com/?target=https%3A//github.com/liucongg/ChatGLM-Finetuning)\n## update-2023.04.18 **增加文本生成任务评测**\n## **ChatGLM-6B模型微调**\n模型越大对[显卡](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%98%BE%E5%8D%A1&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmmL7ljaEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.g4qsqP2-BONGCTtJwXt-G4PMQ4KnwM-GN5jEJF-1S18&zhida_source=entity)的要求越高，目前主流对大模型进行微调方法有三种：[Freeze方法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Freeze%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJGcmVlemXmlrnms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.A4wJCUQe6HfMV-ZWJLM5XMNdfBKFUvMRJk5eoqj7ZC0&zhida_source=entity)、P-Tuning方法和[Lora方法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Lora%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJMb3Jh5pa55rOVIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.BVMSiIU35v84Js5JTFEs2Kz3JvfxwKyO6QL6PCHbhfs&zhida_source=entity)。笔者也通过这三种方法，在信息抽取任务上，对ChatGLM-6B大模型进行模型微调。为了防止大模型的数据泄露，采用一个领域比赛数据集-[汽车工业故障模式关系抽取](https://link.zhihu.com/?target=https%3A//www.datafountain.cn/competitions/584)，随机抽取50条作为测试集。\n详细代码见上面的GitHub链接，并且也被ChatGLM官方收录。 \n![](../images/004.jpg)\n### **Freeze方法**\nFreeze方法，即[参数冻结](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%8F%82%E6%95%B0%E5%86%BB%E7%BB%93&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlj4LmlbDlhrvnu5MiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.34ii6cjSzzY0UrCHFyezXeUjm9mZxovMj85gF4xnGto&zhida_source=entity)，对原始模型部分参数进行冻结操作，仅训练部分参数，以达到在单卡或不进行TP或PP操作，就可以对大模型进行训练。\n微调代码，见finetuning_freeze.py，核心部分如下：\n\n```\nfor name, param in model.named_parameters():\n if not any(nd in name for nd in [\"layers.27\", \"layers.26\", \"layers.25\", \"layers.24\", \"layers.23\"]):\n param.requires_grad = False\n```\n\n针对模型不同层进行修改，可以自行修改。训练代码均采用[DeepSpeed](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=DeepSpeed&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJEZWVwU3BlZWQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.ilWMF-Dn7dgpcJGSjivyvS0Y0t007P-KbAY3eV0By6Y&zhida_source=entity)进行训练，可设置参数包含train_path、model_dir、num_train_epochs、train_batch_size、[gradient_accumulation_steps](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=gradient_accumulation_steps&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJncmFkaWVudF9hY2N1bXVsYXRpb25fc3RlcHMiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.V9xhKiv9aAW_ZmYIra4QBIspQdV1RRGGDj-KpBppFWs&zhida_source=entity)、output_dir、prompt_text等，可根据自己的任务配置。\n\n```\nCUDA_VISIBLE_DEVICES=0 deepspeed finetuning_freeze.py --num_train_epochs 5 --train_batch_size 2\n```\n\n[三元组](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%B8%89%E5%85%83%E7%BB%84&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkuInlhYPnu4QiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Ck0Bf1T78CEl4zmqer4HI_gnl5yIvaPKJ5y-bDU7di0&zhida_source=entity)抽取的[推理代码](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%8E%A8%E7%90%86%E4%BB%A3%E7%A0%81&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmjqjnkIbku6PnoIEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.dF70VbuJ8JDs-_m3qj7jM9GjmatTRwQTvoZqQqMWisk&zhida_source=entity)，见predict_freeze.py，其他任务可以根据自己的评价标准进行推理预测。\n### **PT方法**\nPT方法，即P-Tuning方法，参考[ChatGLM官方代码](https://link.zhihu.com/?target=https%3A//github.com/THUDM/ChatGLM-6B/blob/main/ptuning/README.md) ，是一种针对于大模型的soft-prompt方法。 \n![](../images/005.jpg)\n * [P-Tuning](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2103.10385)，仅对大模型的Embedding加入新的参数。\n * [P-Tuning-V2](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2110.07602)，将大模型的Embedding和每一层前都加上新的参数。\n\n\n微调代码，见finetuning_pt.py，核心部分如下：\n\n```\nconfig = ChatGLMConfig.from_pretrained(args.model_dir)\nconfig.pre_seq_len = args.pre_seq_len\nconfig.prefix_projection = args.prefix_projection\n\nmodel = ChatGLMForConditionalGeneration.from_pretrained(args.model_dir, config=config)\n\nfor name, param in model.named_parameters[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=2&q=named_parameters&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJuYW1lZF9wYXJhbWV0ZXJzIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.Ug45pcpNHAVxfI15ktBqqmAxHpZ3hxB3Y8RA3r5d5iE&zhida_source=entity)():\n if not any(nd in name for nd in [\"prefix_encoder\"]):\n param.requires_grad = False\n```\n\n当[prefix_projection](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=3&q=prefix_projection&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJwcmVmaXhfcHJvamVjdGlvbiIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjIyNjEwNTg3MywiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjozLCJ6ZF90b2tlbiI6bnVsbH0.zbf-2u8C6r689qy9HmAbLMhSx0qNvIy6PMBFNFXbxIU&zhida_source=entity)为True时，为P-Tuning-V2方法，在大模型的Embedding和每一层前都加上新的参数；为False时，为P-Tuning方法，仅在大模型的Embedding上新的参数。\n可设置参数包含train_path、model_dir、num_train_epochs、train_batch_size、gradient_accumulation_steps、output_dir、prompt_text、pre_seq_len、prompt_text等， 可根据自己的任务配置。\n\n```\nCUDA_VISIBLE_DEVICES=0 deepspeed finetuning_pt.py --num_train_epochs 5 --train_batch_size 2 --pre_seq_len 16\n```\n\n三元组抽取的推理代码，见predict_pt.py，其他任务可以根据自己的评价标准进行推理预测。\n### **Lora方法**\nLora方法，即在大型语言模型上对指定参数增加额外的[低秩矩阵](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%BD%8E%E7%A7%A9%E7%9F%A9%E9%98%B5&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkvY7np6nnn6npmLUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Ov4N1t7fgwDtcjeDBqTM__ZGhV693r-2AMpVceEqjSA&zhida_source=entity)，并在模型训练过程中，仅训练而外增加的参数。当“秩值”远小于原始参数维度时，新增的低秩矩阵参数量很小，达到仅训练很小的参数，就能获取较好的结果。 \n![](../images/006.jpg)\n * Lora论文：[Paper](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2106.09685)\n * 官方代码：[Github](https://link.zhihu.com/?target=https%3A//github.com/microsoft/LoRA)\n * HuggingFace封装的peft库：[Github](https://link.zhihu.com/?target=https%3A//github.com/huggingface/peft)\n\n\n微调代码，见finetuning_lora.py，核心部分如下：\n\n```\nmodel = ChatGLMForConditionalGeneration.from_pretrained(args.model_dir)\nconfig = LoraConfig(r=args.lora_r,\n lora_alpha[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=lora_alpha&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJsb3JhX2FscGhhIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Ns3tttuwY6f77vbrhvwUQJub2T9-eqliRyXxYb0P3Gw&zhida_source=entity)=32,\n target_modules=[\"query_key_value\"],\n lora_dropout=0.1,\n bias=\"none\",\n task_type=\"CAUSAL_LM\",\n inference_mode=False,\n )\n\nmodel = get_peft_model[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=peft_model&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJwZWZ0X21vZGVsIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.usI5RIgeF7S8wLJomUWgRCifrtAbNpdBOTWeuYBRaKk&zhida_source=entity)(model, config)\n```\n\n可设置参数包含train_path、model_dir、num_train_epochs、train_batch_size、gradient_accumulation_steps、output_dir、prompt_text、lora_r等，可根据自己的任务配置。\n\n```\nCUDA_VISIBLE_DEVICES=0 deepspeed finetuning_lora.py --num_train_epochs 5 --train_batch_size 2 --lora_r 8\n```\n\n三元组抽取的推理代码，见predict_lora.py，其他任务可以根据自己的评价标准进行推理预测。\n注意：对于结果需要保持一致的任务(即关掉dropout，解码关掉do_sample)，需要保存模型的adapter_config.json文件中，inference_mode参数修改成false，并将模型执行model.eval()操作。 主要原因是chatglm模型代码中，没有采用Conv1D函数。\n## **三元组抽取实验结果**\n * 模型训练时，最大长度为768，Batch大小为2，训练轮数为5，fp16训练，采用DeepSpeed的Zero-1训练；\n * PT为官方的P-Tuning V2[训练方法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%AE%AD%E7%BB%83%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLorq3nu4Pmlrnms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.6rN-pl1ivWQHYI94TBQ8UO9tm4iUCsWdKUpWhjKoMp4&zhida_source=entity)，PT-Only-Embedding表示仅对Embedding进行soft-prompt，Freeze仅训练模型后五层参数，Lora采用低秩矩阵方法训练，秩为8；\n * 由于之前训练PT在48G-A40显卡上会出现OOM，因此之前进行PT实验时对模型开启了gradient_checkpointing_enable，使得模型显存占用变小，但训练时长增加。\n * 训练示例：\n\n\n\n```\nprompt_text：你现在是一个信息抽取模型，请你帮我抽取出关系内容为\\\"性能故障\\\", \\\"部件故障\\\", \\\"组成\\\"和 \\\"检测工具\\\"的相关三元组，三元组内部用\\\"_\\\"连接，三元组之间用\\\\n分割。文本：\n输入：故障现象：发动机水温高，风扇始终是低速转动，高速档不工作，开空调尤其如此。\n输出：发动机_部件故障_水温高\\n风扇_部件故障_低速转动\n```\n\n时间换空间，可用很好的解决显卡的资源问题，简单玩玩还可以，如果想要模型达到最优效果或可用快速看到效果，还不如租张A100卡，快速实验，推理阶段再用自己的小破卡。\n下面实验结果均是在租的80G-A100上进行的实验，与Github里用的A40的实验结果会有些差异，主要在训练时长(纯[训练速度](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%AE%AD%E7%BB%83%E9%80%9F%E5%BA%A6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLorq3nu4PpgJ_luqYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.t-AmRGbze4gLsjexQKbxq-v0D80rV6tiq80Ul2u100s&zhida_source=entity)，剔除模型保存的时间)。说实话，真的要训练一个大模型，多个A100是必不可少的，可以减少很多模型并行的操作，效果上也更好把控一些。 \n| 微调方法 | PT-Only-Embedding | PT | Freeze | Lora | \n| --- | --- | --- | --- | --- | \n| 显卡占用 | 37G | 56G | 24G | 39G | \n| 总参数 | 6.259B | 7.211B | 6.255B | 6.259B | \n| 可训练参数占比 | 0.0586% | 13.26% | 16.10% | 0.0586% | \n| 训练耗时 | 20min | 52min | 46min | 25min | \n| 测试结果F1 | 0.0 | 0.6283 | 0.5675 | 0.5359 | \n结果分析：\n * 效果为PT>Freeze>Lora>PT-Only-Embedding;\n * 速度为PT-Only-Embedding>Lora>Freeze>PT;\n * **PT-Only-Embedding效果很不理想，发现在训练时，最后的loss仅能收敛到2.几，而其他机制可以收敛到0.几。分析原因为，输出内容形式与原有语言模型任务相差很大，仅增加额外Embedding参数，不足以改变复杂的[下游任务](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%B8%8B%E6%B8%B8%E4%BB%BB%E5%8A%A1&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkuIvmuLjku7vliqEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.689UbNINqmQlBd2lHtDgyjruko4c9-orNeyeGV78czQ&zhida_source=entity)**;\n * PT方法占用[显存](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=2&q=%E6%98%BE%E5%AD%98&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmmL7lrZgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MiwiemRfdG9rZW4iOm51bGx9.CtSACxtivtl_MMRqfYJP8ICzvtt1wIiE6XfN-z-FVyo&zhida_source=entity)更大，因为也增加了很多而外参数;\n * 测试耗时，采用float16进行[模型推理](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmqKHlnovmjqjnkIYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.IULSkFKFuuVYRH-uC3v4vqUEDytRGKvdXVpjMRJ0epQ&zhida_source=entity)，由于其他方法均增加了额外参数，因此其他方法的推理耗时会比Freeze方法要高。当然由于是生成模型，所以生成的长度也会影响耗时;\n * 模型在指定任务上微调之后，并没有丧失原有能力，例如生成“帮我写个[快排算法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%BF%AB%E6%8E%92%E7%AE%97%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlv6vmjpLnrpfms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.UFSJHLHGt7WiEp4XQrrFoxx2NQOZ0cPUhSxivPzdedQ&zhida_source=entity)”，依然可以生成-快排代码;\n * **由于[大模型微调](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=2&q=%E5%A4%A7%E6%A8%A1%E5%9E%8B%E5%BE%AE%E8%B0%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlpKfmqKHlnovlvq7osIMiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MiwiemRfdG9rZW4iOm51bGx9.sWKd6yIQXkr3qGoYR2V231GCQUP3mCdubzL-ITAfzDU&zhida_source=entity)都采用大量instruction进行模型训练，仅采用单一的指令进行微调时，对原来其他的指令影响不大，因此并没导致原来模型的能力丧失**;\n * 上面测试仅代表个人测试结果。\n\n\n很多同学在微调后出现了灾难性遗忘现象，但我这边并没有出现，对“翻译任务”、“代码任务”、“问答任务”进行测试，采用freeze模型，可以用test_forgetting.py进行测试，具体测试效果如下：\n * 翻译任务 \n\n\n![](../images/007.png)\n * 代码任务 \n\n\n![](../images/008.jpg)\n * 问答任务 \n\n\n![](../images/009.png)\n后面会把生成任务、分类任务做完，请持续关注Github，会定期更新。（太忙了，会抓紧时间更新，并且官方代码也在持续更新，如遇到代码代码调不通的情况，请及时联系我，我在github也给出了我的代码版本和模型版本）\n### **文本生成**\n * 为了防止大模型的数据泄露，采用一个“万创杯”[中医药天池大数据竞赛](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%B8%AD%E5%8C%BB%E8%8D%AF%E5%A4%A9%E6%B1%A0%E5%A4%A7%E6%95%B0%E6%8D%AE%E7%AB%9E%E8%B5%9B&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkuK3ljLvoja_lpKnmsaDlpKfmlbDmja7nq57otZsiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.VByrU5xRrBIyvUU6GvrupUs9fCE1XWUA1Dp0b71D_Ws&zhida_source=entity)-[中医文献问题生成挑战](https://link.zhihu.com/?target=https%3A//tianchi.aliyun.com/competition/entrance/531826/introduction)，随机抽取20条作为测试集\n * PT为官方的P-Tuning V2训练方法，PT-Only-Embedding表示仅对Embedding进行soft-prompt，Freeze仅训练模型后五层参数，Lora采用低秩矩阵方法训练，秩为8；\n * 训练示例：\n\n\n\n```\nprompt_text：你现在是一个问题生成模型，请根据下面文档生成一个问题，文档：\n输入：紫色红薯是近年从日本[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%97%A5%E6%9C%AC&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLml6XmnKwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.-32xvuaLamIVhtZxcDg4pWwnBjqrFSTn-_x3IXcorik&zhida_source=entity)引进的新品种红薯，中国农业大学农学与生物技术学院副院长刘庆昌[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%88%98%E5%BA%86%E6%98%8C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLliJjluobmmIwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.ydJi2gCYBnZesDRrNQ_RZgHtG03KbZi7rhBvP-jgtss&zhida_source=entity)指出，紫薯中的花青素[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%8A%B1%E9%9D%92%E7%B4%A0&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLoirHpnZLntKAiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.A9NiU9LOrtSRTRk9QU9KuEKxeXgFk-zPEMaBGceWJxk&zhida_source=entity)具有显著的抗生物[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%8A%97%E7%94%9F%E7%89%A9&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmipfnlJ_niakiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.4qoI7eNOv9Hcu9s09Bthm3fExl2e3FoTa2sVnp8xpmA&zhida_source=entity)氧化作用，在延缓人体衰老方面具有非常好的效果。紫薯中所含赖氨酸[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%B5%96%E6%B0%A8%E9%85%B8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLotZbmsKjphbgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.tJA7ZRuA34EHtSoYnrKL_oeuz00EmmE2wFVqrS4B3Ak&zhida_source=entity)、铜、锰、钾、锌的含量高于一般红薯5-8倍，尤其是抗癌物质碘、硒的含量比其他红薯高出20倍以上，占食物中的第一位。\n输出：紫薯和红薯吃哪个好？\n```\n\n[模型训练](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=4&q=%E6%A8%A1%E5%9E%8B%E8%AE%AD%E7%BB%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmqKHlnovorq3nu4MiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6NCwiemRfdG9rZW4iOm51bGx9.da8PBQ0lNUyi0mwwnoACUdwYNotYbeZk7GhgKpXfZ1k&zhida_source=entity)，以freeze方法为例：\n\n```\nCUDA_VISIBLE_DEVICES=0 nohup deepspeed --master_port 5555 finetuning_freeze.py --train_path \"data/d2q_0.json\" --output_dir \"output_dir_freeze/\" --prompt_text \"你现在是一个问题生成模型，请根据下面文档生成一个问题，文档：\" > log_fz.log 2>&1 &\n```\n\n由于生成模型的内容不能想信息抽取任务一样评价，用现有的BLUE或者[Rouge](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Rouge&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJSb3VnZSIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjIyNjEwNTg3MywiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.hUI4f9AWlnY2-5_5tlVgmWzULKBbbr1KSD6slj7UFiU&zhida_source=entity)来评价也是不合适，因此制定了评分规则。 通过多样性和准确性两个角度判断D2Q模型好坏，每个样本总计5分，共20个样本。\n * 多样性： \n * 问题是否高度相似，每重复一个问题扣0.25分；\n * 问题对应答案是否相同，每有一个重复答案或找不到答案，扣0.25分；\n * 准确性： \n * 问题能否从文档中找到答案，每有一个找不到答案，扣0.25分；\n * 问题内容是否流畅，每有一个问题不流畅，扣0.25分；\n * 问题内容是否有害，每有一个有害，扣0.25分；\n\n\n测试数据见d2q_result_data/，测试代码见predict_d2q.py \n| 微调方法 | 原始模型 | PT-Only-Embedding | PT | Freeze | Lora | \n| --- | --- | --- | --- | --- | --- | \n| 分数 | 51.75 | 73.75 | 87.75 | 79.25 | 86.75 | \n## **中文开源大模型 &项目**\n虽然出来很多大模型，但Open的&中文可直接使用的并不多，下面对中文开源大模型、数据集和项目进行一下汇总。 \n![](../images/010.jpg)\n### **中文开源大模型**\n直接可微调，无需指令[增量训练](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%A2%9E%E9%87%8F%E8%AE%AD%E7%BB%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlop7ph4_orq3nu4MiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.aszCUw9y4N-vuDpdPwOapcfp38bMXsfp98OSkhY9-GA&zhida_source=entity)：\n * ChatGLM-6B：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/THUDM/chatglm-6b)\n * ChatYuan-large-v2：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/ClueAI/ChatYuan-large-v2)\n\n\n原始模型多语言or英文，需要中文指令数据集增量训练：\n * BloomZ：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/bigscience/bloomz)\n * LLama：[模型地址](https://link.zhihu.com/?target=https%3A//github.com/facebookresearch/llama)\n * Flan-T5：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/google/flan-t5-xxl)\n * OPT：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/facebook/opt-66b)\n\n\n### **中文开源指令数据**\n下面中文[指令集](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%8C%87%E4%BB%A4%E9%9B%86&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmjIfku6Tpm4YiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Je9Z6VreHMwaGkowyOWtcGDslowmn7Q6r8OoufT-3fs&zhida_source=entity)，大多数从Alpaca翻译而来，请看下面项目中data目录。目前通过ChatGPT或者GPT4作为廉价标注工为自己的数据进行数据标注一个不错的思路。\n * [1]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/LC1332/Chinese-alpaca-lora)\n * [2]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/hikariming/alpaca_chinese_dataset)\n * [3]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/carbonz0/alpaca-chinese-dataset)\n * [4]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)\n * [5]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/LianjiaTech/BELLE)\n * [6]：[数据地址](https://link.zhihu.com/?target=https%3A//huggingface.co/datasets/JosephusCheung/GuanacoDataset)\n\n\n### **开源项目**\n总结下面较火的开源项目：\n * BELLE：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/LianjiaTech/BELLE)\n * ChatGLM：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/THUDM/ChatGLM-6B)\n * Luotuo-Chinese-LLM：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/LC1332/Luotuo-Chinese-LLM)\n * stanford_alpaca：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/tatsu-lab/stanford_alpaca)\n\n\n## **总结**\n目前各大厂的大模型陆陆续续放出，堪称百家争鸣！个人玩家也是全面拥抱，想尽一切办法来训练微调大模型。只愿大家以后可以实现“大模型”自由。愿再无“[model-as-a-service](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=model-as-a-service&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJtb2RlbC1hcy1hLXNlcnZpY2UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.6bAH0u0Hx6-tNterCdAl8vv8rlEpNDWKi_1f4shS2oI&zhida_source=entity)”。\n* * *\n**欢迎多多转发，点赞，关注，有问题的朋友也欢迎加我微信** 「logCong」**私聊，交个朋友吧，一起学习，一起进步。**\n**我们的口号是“生命不止，学习不停”。**\n## **往期回顾**\n[刘聪NLP：ChatGPT-所见、所闻、所感](https://zhuanlan.zhihu.com/p/605331104)\n[刘聪NLP：AAAI2023 | 基于统一语义匹配的通用信息抽取框架-USM](https://zhuanlan.zhihu.com/p/598882471)\n[刘聪NLP：CC-Riddle：汉字谜语问答数据集](https://zhuanlan.zhihu.com/p/537552626)\n[刘聪NLP：中文NER数据集整理](https://zhuanlan.zhihu.com/p/529541521)\n[刘聪NLP：ACL2022 | DCSR：一种面向开放域段落检索的句子感知的对比学习方法](https://zhuanlan.zhihu.com/p/527366495)\n[刘聪NLP：ACL2022 | NoisyTune：微调前加入少量噪音可能会有意想不到的效果](https://zhuanlan.zhihu.com/p/523865674)\n[刘聪NLP：ACL2022|DictBERT：通过低频词典增强预训练模型表征方法](https://zhuanlan.zhihu.com/p/523551515)\n[刘聪NLP：ACL2022论文分类汇总-Prompt、句子表征、检索排序&摘要](https://zhuanlan.zhihu.com/p/519943388)\n[刘聪NLP：总结|Prompt在NER场景的应用](https://zhuanlan.zhihu.com/p/518146549)\n[刘聪NLP：PolyLoss：一种将分类损失函数加入泰勒展开式的损失函数](https://zhuanlan.zhihu.com/p/510626670)\n[刘聪NLP：PERT：一种基于乱序语言模型的预训练模型](https://zhuanlan.zhihu.com/p/509647368)\n[刘聪NLP：DiffCSE：结合句子间差异的无监督句子嵌入对比学习方法](https://zhuanlan.zhihu.com/p/507171467)\n[刘聪NLP：SIGIR2022论文筛选](https://zhuanlan.zhihu.com/p/506005620)\n[刘聪NLP：PairSCL：句子对级别的有监督对比学习方法](https://zhuanlan.zhihu.com/p/463949437)\n[刘聪NLP：COLD：中文冒犯性语言检测数据集](https://zhuanlan.zhihu.com/p/463455280)\n[刘聪NLP：SNCSE：一种基于软负例的无监督句向量对比学习方法](https://zhuanlan.zhihu.com/p/463142612)\n[刘聪NLP：CPT模型：一种中文兼顾NLU和NLG的非平衡预训练语言模型](https://zhuanlan.zhihu.com/p/456553852)\n[刘聪NLP：SimCSE论文精读](https://zhuanlan.zhihu.com/p/452761704)\n[刘聪NLP：常用预训练语言模型（PTMs）总结](https://zhuanlan.zhihu.com/p/406512290)\n[刘聪NLP：ACL2021论文之ChineseBERT：融合字形与拼音信息的中文预训练模型](https://zhuanlan.zhihu.com/p/393617564)\n[刘聪NLP：Sentence-Bert论文笔记](https://zhuanlan.zhihu.com/p/113133510)\n[刘聪NLP：MacBERT：MLM as correction BERT](https://zhuanlan.zhihu.com/p/250595837)\n[刘聪NLP：智能扩充机器人的“标准问”库之Query生成](https://zhuanlan.zhihu.com/p/149429784)\n[刘聪NLP：短文本相似度算法研究](https://zhuanlan.zhihu.com/p/111414376)\n编辑于 2023-06-30 12:44・江苏\n[LLM（大型语言模型）](https://www.zhihu.com/topic/26797383)\n[自然语言处理](https://www.zhihu.com/topic/19560026)\n[人工智能](https://www.zhihu.com/topic/19551275)\n​赞同 1731​​167 条评论​4044 ​318 \n​分享\n​申请转载​\n​\n![](../images/011.jpeg)\n未登录用户\n \n\n167 条评论\n默认\n最新\n[![刘聪NLP](../images/023.jpg)](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n[刘聪NLP](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n作者​![](../images/body-2dd3e5c360100c555650.png)\n更新了文本生成的实验![\\[吃瓜\\]](https://pic4.zhimg.com/v2-74ecc4b114fce67b6b42b7f602c3b1d6.png)\n2023-04-18 · 江苏 · 作者置顶\n​回复​8\n[![徐晓龙](../images/014.jpg)](https://www.zhihu.com/people/6785657c61e5032307fce9e7f7c5ad9e)\n[徐晓龙](https://www.zhihu.com/people/6785657c61e5032307fce9e7f7c5ad9e)\n你这个里面没看到没微调前的模型效果，有这块的数据嘛\n2023-04-14 · 安徽\n​回复​8\n[![Szymon](../images/015.jpg)](https://www.zhihu.com/people/4eb0e2b49dd64734145f896f3303b982)\n[Szymon](https://www.zhihu.com/people/4eb0e2b49dd64734145f896f3303b982)\n这里有条挺重要的，专项训练过后是否影响原先的各种性能。\n2023-04-20 · 四川\n​回复​5\n展开其他 4 条回复​\n[![半糖](../images/019.jpg)](https://www.zhihu.com/people/a55028acdb8c97237a8fba9940a2ae95)\n[半糖](https://www.zhihu.com/people/a55028acdb8c97237a8fba9940a2ae95)\n大佬能做下文本分类的微调吗～最近在做感觉效果很差，不知道哪里出问题了55\n2023-05-17 · 福建\n​回复​5\n展开其他 5 条回复​\n[![揽睿星舟](../images/017.jpg)](https://www.zhihu.com/org/9bdb83ab64dc4b31656f2c14f8f48eff)\n[揽睿星舟](https://www.zhihu.com/org/9bdb83ab64dc4b31656f2c14f8f48eff)\n妈妈！我上电视了！![\\[爱\\]](https://pic1.zhimg.com/v2-0942128ebfe78f000e84339fbb745611.png)\n2023-04-11 · 上海\n​回复​4\n[![刘聪NLP](../images/018.jpg)](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n[刘聪NLP](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n作者​![](../images/body-2dd3e5c360100c555650.png)\n![\\[酷\\]](https://pic4.zhimg.com/v2-c96dd18b15beb196b2daba95d26d9b1c.png)![\\[酷\\]](https://pic4.zhimg.com/v2-c96dd18b15beb196b2daba95d26d9b1c.png)\n2023-04-11 · 江苏\n​回复​喜欢\n展开其他 1 条回复​\n[![NyChatAI](../images/019.jpg)](https://www.zhihu.com/people/4ac9362ebdbb221ab13ebaa9b0180a75)\n[NyChatAI](https://www.zhihu.com/people/4ac9362ebdbb221ab13ebaa9b0180a75)\n为什么指令微调的输入是input+label：输出是input的长度*100；能采用input，输出直接是label吗？\n2023-07-24 · 广东\n​回复​2\n[![Lemon](../images/020.jpg)](https://www.zhihu.com/people/b2cab50cbd39c8138c9a7f0a06f341a8)\n[Lemon](https://www.zhihu.com/people/b2cab50cbd39c8138c9a7f0a06f341a8)\n楼主好，想请教一下，关系抽取任务没有影响模型原来的自然语言理解能力，那生成任务有影响嘛？感觉关系抽取任务可能在之前预训练的时候，用到了类似的数据或者指令，所以没有对原始模型产生太大的影响。如果是一个完全特定领域的生成任务，会有影响嘛？\n2023-05-19 · 北京\n​回复​2\n[![wyl](../images/021.jpg)](https://www.zhihu.com/people/5cc2745fb4c468616a02e712e520c97d)\n[wyl](https://www.zhihu.com/people/5cc2745fb4c468616a02e712e520c97d)\n当prefix_projection为True时，为P-Tuning-V2方法，在大模型的Embedding和每一层前都加上新的参数；为False时，为P-Tuning方法，仅在大模型的Embedding上新的参数。这个有错误吧？True代表的是使用了重参数吧？\n2023-05-18 · 北京\n​回复​2\n展开其他 4 条回复​\n[![王二叔](../images/022.jpg)](https://www.zhihu.com/people/e630a433d5d02d31e3182c6507612fea)\n[王二叔](https://www.zhihu.com/people/e630a433d5d02d31e3182c6507612fea)\n感谢分享 \n2023-04-12 · 北京\n​回复​1\n[![刘聪NLP](../images/023.jpg)](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n[刘聪NLP](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n作者​![](../images/body-2dd3e5c360100c555650.png)\n![\\[吃瓜\\]](https://pic4.zhimg.com/v2-74ecc4b114fce67b6b42b7f602c3b1d6.png)\n2023-04-12 · 江苏\n​回复​喜欢\n点击查看全部评论\n![](../images/011.jpeg)\n未登录用户\n \n\n关于作者\n[![刘聪NLP](../images/024.jpg)](https://www.zhihu.com/people/LiuCongNLP)\n[刘聪NLP](https://www.zhihu.com/people/LiuCongNLP)\nLLM is all you need！\n[​](https://zhuanlan.zhihu.com/p/96956163)\n云问科技 算法工程师\n[ 回答 **505** ](https://www.zhihu.com/people/LiuCongNLP/answers)[ 文章 **205** ](https://www.zhihu.com/people/LiuCongNLP/posts)[ 关注者 **35,939** ](https://www.zhihu.com/people/LiuCongNLP/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[江歌妈妈发长文尘埃终将落定706 万](https://www.zhihu.com/search?q=%E6%B1%9F%E6%AD%8C%E5%A6%88%E5%A6%88%E5%8F%91%E9%95%BF%E6%96%87%E5%B0%98%E5%9F%83%E7%BB%88%E5%B0%86%E8%90%BD%E5%AE%9A&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[中国男足 0-5 巴勒斯坦564 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[原央视主持人阿丘回应被通报422 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[华为与赛力斯达成新五年合作411 万](https://www.zhihu.com/search?q=%E5%8D%8E%E4%B8%BA%E4%B8%8E%E8%B5%9B%E5%8A%9B%E6%96%AF%E8%BE%BE%E6%88%90%E6%96%B0%E4%BA%94%E5%B9%B4%E5%90%88%E4%BD%9C&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[迪拜航空确认航班发生事故400 万](https://www.zhihu.com/search?q=%E8%BF%AA%E6%8B%9C%E8%88%AA%E7%A9%BA%E7%A1%AE%E8%AE%A4%E8%88%AA%E7%8F%AD%E5%8F%91%E7%94%9F%E4%BA%8B%E6%95%85&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[东航回应网传空姐跪地道歉327 万](https://www.zhihu.com/search?q=%E4%B8%9C%E8%88%AA%E5%9B%9E%E5%BA%94%E7%BD%91%E4%BC%A0%E7%A9%BA%E5%A7%90%E8%B7%AA%E5%9C%B0%E9%81%93%E6%AD%89&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n[王楚钦：不知道为什么就是感觉累320 万](https://www.zhihu.com/search?q=%E7%8E%8B%E6%A5%9A%E9%92%A6%EF%BC%9A%E4%B8%8D%E7%9F%A5%E9%81%93%E4%B8%BA%E4%BB%80%E4%B9%88%E5%B0%B1%E6%98%AF%E6%84%9F%E8%A7%89%E7%B4%AF&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[江苏高考作文《衬衫的价格为 9 磅 15 便士》爆火268 万](https://www.zhihu.com/search?q=%E6%B1%9F%E8%8B%8F%E9%AB%98%E8%80%83%E4%BD%9C%E6%96%87%E3%80%8A%E8%A1%AC%E8%A1%AB%E7%9A%84%E4%BB%B7%E6%A0%BC%E4%B8%BA+9+%E7%A3%85+15+%E4%BE%BF%E5%A3%AB%E3%80%8B%E7%88%86%E7%81%AB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n[女装网店开始用防拆带了265 万](https://www.zhihu.com/search?q=%E5%A5%B3%E8%A3%85%E7%BD%91%E5%BA%97%E5%BC%80%E5%A7%8B%E7%94%A8%E9%98%B2%E6%8B%86%E5%B8%A6%E4%BA%86&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n[孕妇骑车别车被司机踹翻258 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n### 推荐阅读\n# [llm微调记录 03 微调方法LlamaFactory 的微调方法有3种： Full： \\theta_{new}=\\theta_{old}+\\Delta\\theta Lora: \\theta_{new}=\\theta_{old}+B\\cdot A^T Freeze:冻结层： \\theta_{new}=\\theta_{old} 训练层… 谁知道呢 ](https://zhuanlan.zhihu.com/p/1907910324352456379)[ ![模型应用-LLM微调](../images/026.jpg)模型应用-LLM微调 Dev-Z...发表于模型方法 ](https://zhuanlan.zhihu.com/p/676490419)[ ![LLM微调实践和总结](../images/028.jpg)LLM微调实践和总结 yuwei1991 ](https://zhuanlan.zhihu.com/p/684763523)[ ![大模型学习笔记（四） 更好的LLM：微调、提示词工程与对齐](../images/030.jpg)大模型学习笔记（四） 更好的LLM：微调、提示词工程与对齐 水滴油穿 ](https://zhuanlan.zhihu.com/p/714988018) \n\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](../images/body-0f9788d006ab0f4da762.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n", "metadata": { "title": "大模型LLM-微调经验分享&总结 - 知乎", "description": "写在前面大型语言模型横行，之前非常焦虑，现在全面拥抱。目前也有很多开源项目进行大模型微调等，笔者也做了一阵子大模型了，特此来介绍一下ChatGLM-6B模型微调经验，并汇总了一下目前开源项目&数据。笔者与…", "keywords": "LLM（大型语言模型）,自然语言处理,人工智能", "author": null, "og:title": "大模型LLM-微调经验分享&总结", "og:url": "https://zhuanlan.zhihu.com/p/620885226", "og:description": "写在前面大型语言模型横行，之前非常焦虑，现在全面拥抱。目前也有很多开源项目进行大模型微调等，笔者也做了一阵子大模型了，特此来介绍一下ChatGLM-6B模型微调经验，并汇总了一下目前开源项目&数据。笔者与…", "og:image": "https://picx.zhimg.com/v2-63e76158855321ccb104fc9b3402005d_720w.jpg?source=172ae18b", "og:type": "article", "og:site_name": "知乎专栏" }, "paywall": { "listed_paid": false, "detected": false, "marker": "", "public_only": false, "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙" }, "crawl": { "success": true, "status_code": 200, "error": "" }, "code_blocks": [ { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_000.txt", "chars": 185 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_001.txt", "chars": 96 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_002.txt", "chars": 812 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_003.txt", "chars": 109 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_004.txt", "chars": 1271 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_005.txt", "chars": 105 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_006.txt", "chars": 181 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_007.txt", "chars": 2409 }, { "language": "txt", "file": "attachments/e0eda9767a9f/code/block_008.txt", "chars": 211 } ], "images": [ { "status": "http_403", "url": "https://zhuanlan.zhihu.com/p/620885226", "src": "https://zhuanlan.zhihu.com/p/620885226", "alt": "", "downloaded": false, "discovery": "saved_markdown", "retried_at": "2026-10-20T18:08:42Z" } ], "audio": [], "videos": [], "links": [ { "type": "internal", "href": "https://www.zhihu.com" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Ffollow" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2F" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fhot" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fcolumn-square" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fring-feeds" }, { "type": "internal", "href": "https://www.zhihu.com/project-square" }, { "type": "internal", "href": "https://www.zhihu.com/fiore/h5/vip-web" }, { "type": "internal", "href": "https://zhida.zhihu.com/" }, { "type": "internal", "href": "https://www.zhihu.com/column/c_1625207909599617024" }, { "type": "internal", "href": "https://www.zhihu.com/people/LiuCongNLP" }, { "type": "internal", "href": "https://www.zhihu.com/question/510340037" }, { "type": "internal", "href": "https://www.zhihu.com/column/c_1332763253097734144" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E7%81%BE%E9%9A%BE%E6%80%A7%E9%81%97%E5%BF%98&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLngb7pmr7mgKfpgZflv5giLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.m2m3IzZUhVIJqj3GlQeWAfn-TCtYst79-om2wR_2qgk&zhida_source=entity" }, { "type": "internal", "href": "https://link.zhihu.com/?target=https%3A%2F%2Fgithub.com%2Fliucongg%2FChatGLM-Finetuning" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%98%BE%E5%8D%A1&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmmL7ljaEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.g4qsqP2-BONGCTtJwXt-G4PMQ4KnwM-GN5jEJF-1S18&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Freeze%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJGcmVlemXmlrnms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.A4wJCUQe6HfMV-ZWJLM5XMNdfBKFUvMRJk5eoqj7ZC0&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Lora%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJMb3Jh5pa55rOVIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.BVMSiIU35v84Js5JTFEs2Kz3JvfxwKyO6QL6PCHbhfs&zhida_source=entity" }, { "type": "internal", "href": "https://link.zhihu.com/?target=https%3A%2F%2Fwww.datafountain.cn%2Fcompetitions%2F584" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%8F%82%E6%95%B0%E5%86%BB%E7%BB%93&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlj4LmlbDlhrvnu5MiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.34ii6cjSzzY0UrCHFyezXeUjm9mZxovMj85gF4xnGto&zhida_source=entity" }, { "type": "external", "href": "javascript:void(0)" } ], "repair_version": 2, "repaired_at": "2026-10-20T17:33:58Z", "source_url": "https://zhuanlan.zhihu.com/p/620885226", "delivery_status": "unverified", "discovery_status": "http_403", "missing_reason": "源页面未重新核验:http_403；已按保留正文补采图片", "repair_error": "", "actual_video_files": [], "expected_saved_body_images": 1, "image_discovery": "saved_markdown", "source_body_markdown": "[![](https://zhuanlan.zhihu.com/p/620885226)](javascript:void\\(0\\))\n![ZhiHu logo](https://static.zhihu.com/heifetz/assets/wechat-share-logo.39ea9ecd.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n大模型LLM-微调经验分享&总结\n首发于[AIGC](https://www.zhihu.com/column/c_1625207909599617024)\n切换模式\n![大模型LLM-微调经验分享&总结](https://picx.zhimg.com/70/v2-63e76158855321ccb104fc9b3402005d_1440w.image?source=172ae18b&biz_tag=Post)\n# 大模型LLM-微调经验分享&总结\n[![刘聪NLP](https://pic1.zhimg.com/v2-53ee42573c43af1ebc16953b5dbd8768_l.jpg?source=32738c0c&needBackground=1)](https://www.zhihu.com/people/LiuCongNLP)\n[刘聪NLP](https://www.zhihu.com/people/LiuCongNLP)\n[​![](https://picx.zhimg.com/v2-27bfcba90e66db79ce8768ab807e017e_l.png?source=32738c0c)](https://www.zhihu.com/question/510340037)\n新知答主\n[收录于 · 人工智能研究](https://www.zhihu.com/column/c_1332763253097734144)\n1731 人赞同了该文章\n## **写在前面**\n大型语言模型横行，之前非常焦虑，现在全面拥抱。目前也有很多开源项目进行大模型微调等，笔者也做了一阵子大模型了，特此来介绍一下ChatGLM-6B模型微调经验，并汇总了一下目前开源项目&数据。笔者与很多人微调结论不同，本人在采用单指令上进行模型微调，发现模型微调之后，**「并没有出现[灾难性遗忘](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E7%81%BE%E9%9A%BE%E6%80%A7%E9%81%97%E5%BF%98&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLngb7pmr7mgKfpgZflv5giLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.m2m3IzZUhVIJqj3GlQeWAfn-TCtYst79-om2wR_2qgk&zhida_source=entity)现象」**。\n[](https://link.zhihu.com/?target=https%3A//github.com/liucongg/ChatGLM-Finetuning)\n## update-2023.04.18 **增加文本生成任务评测**\n## **ChatGLM-6B模型微调**\n模型越大对[显卡](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%98%BE%E5%8D%A1&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmmL7ljaEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.g4qsqP2-BONGCTtJwXt-G4PMQ4KnwM-GN5jEJF-1S18&zhida_source=entity)的要求越高，目前主流对大模型进行微调方法有三种：[Freeze方法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Freeze%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJGcmVlemXmlrnms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.A4wJCUQe6HfMV-ZWJLM5XMNdfBKFUvMRJk5eoqj7ZC0&zhida_source=entity)、P-Tuning方法和[Lora方法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Lora%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJMb3Jh5pa55rOVIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.BVMSiIU35v84Js5JTFEs2Kz3JvfxwKyO6QL6PCHbhfs&zhida_source=entity)。笔者也通过这三种方法，在信息抽取任务上，对ChatGLM-6B大模型进行模型微调。为了防止大模型的数据泄露，采用一个领域比赛数据集-[汽车工业故障模式关系抽取](https://link.zhihu.com/?target=https%3A//www.datafountain.cn/competitions/584)，随机抽取50条作为测试集。\n详细代码见上面的GitHub链接，并且也被ChatGLM官方收录。 \n![](https://pic2.zhimg.com/v2-f048824d8732efb7bc97b27996d88f03_1440w.jpg)\n### **Freeze方法**\nFreeze方法，即[参数冻结](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%8F%82%E6%95%B0%E5%86%BB%E7%BB%93&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlj4LmlbDlhrvnu5MiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.34ii6cjSzzY0UrCHFyezXeUjm9mZxovMj85gF4xnGto&zhida_source=entity)，对原始模型部分参数进行冻结操作，仅训练部分参数，以达到在单卡或不进行TP或PP操作，就可以对大模型进行训练。\n微调代码，见finetuning_freeze.py，核心部分如下：\n\n```\nfor name, param in model.named_parameters():\n if not any(nd in name for nd in [\"layers.27\", \"layers.26\", \"layers.25\", \"layers.24\", \"layers.23\"]):\n param.requires_grad = False\n```\n\n针对模型不同层进行修改，可以自行修改。训练代码均采用[DeepSpeed](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=DeepSpeed&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJEZWVwU3BlZWQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.ilWMF-Dn7dgpcJGSjivyvS0Y0t007P-KbAY3eV0By6Y&zhida_source=entity)进行训练，可设置参数包含train_path、model_dir、num_train_epochs、train_batch_size、[gradient_accumulation_steps](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=gradient_accumulation_steps&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJncmFkaWVudF9hY2N1bXVsYXRpb25fc3RlcHMiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.V9xhKiv9aAW_ZmYIra4QBIspQdV1RRGGDj-KpBppFWs&zhida_source=entity)、output_dir、prompt_text等，可根据自己的任务配置。\n\n```\nCUDA_VISIBLE_DEVICES=0 deepspeed finetuning_freeze.py --num_train_epochs 5 --train_batch_size 2\n```\n\n[三元组](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%B8%89%E5%85%83%E7%BB%84&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkuInlhYPnu4QiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Ck0Bf1T78CEl4zmqer4HI_gnl5yIvaPKJ5y-bDU7di0&zhida_source=entity)抽取的[推理代码](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%8E%A8%E7%90%86%E4%BB%A3%E7%A0%81&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmjqjnkIbku6PnoIEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.dF70VbuJ8JDs-_m3qj7jM9GjmatTRwQTvoZqQqMWisk&zhida_source=entity)，见predict_freeze.py，其他任务可以根据自己的评价标准进行推理预测。\n### **PT方法**\nPT方法，即P-Tuning方法，参考[ChatGLM官方代码](https://link.zhihu.com/?target=https%3A//github.com/THUDM/ChatGLM-6B/blob/main/ptuning/README.md) ，是一种针对于大模型的soft-prompt方法。 \n![](https://pic2.zhimg.com/v2-0906021c2338dfe9c60de250c3965a55_1440w.jpg)\n * [P-Tuning](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2103.10385)，仅对大模型的Embedding加入新的参数。\n * [P-Tuning-V2](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2110.07602)，将大模型的Embedding和每一层前都加上新的参数。\n\n\n微调代码，见finetuning_pt.py，核心部分如下：\n\n```\nconfig = ChatGLMConfig.from_pretrained(args.model_dir)\nconfig.pre_seq_len = args.pre_seq_len\nconfig.prefix_projection = args.prefix_projection\n\nmodel = ChatGLMForConditionalGeneration.from_pretrained(args.model_dir, config=config)\n\nfor name, param in model.named_parameters[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=2&q=named_parameters&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJuYW1lZF9wYXJhbWV0ZXJzIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.Ug45pcpNHAVxfI15ktBqqmAxHpZ3hxB3Y8RA3r5d5iE&zhida_source=entity)():\n if not any(nd in name for nd in [\"prefix_encoder\"]):\n param.requires_grad = False\n```\n\n当[prefix_projection](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=3&q=prefix_projection&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJwcmVmaXhfcHJvamVjdGlvbiIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjIyNjEwNTg3MywiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjozLCJ6ZF90b2tlbiI6bnVsbH0.zbf-2u8C6r689qy9HmAbLMhSx0qNvIy6PMBFNFXbxIU&zhida_source=entity)为True时，为P-Tuning-V2方法，在大模型的Embedding和每一层前都加上新的参数；为False时，为P-Tuning方法，仅在大模型的Embedding上新的参数。\n可设置参数包含train_path、model_dir、num_train_epochs、train_batch_size、gradient_accumulation_steps、output_dir、prompt_text、pre_seq_len、prompt_text等， 可根据自己的任务配置。\n\n```\nCUDA_VISIBLE_DEVICES=0 deepspeed finetuning_pt.py --num_train_epochs 5 --train_batch_size 2 --pre_seq_len 16\n```\n\n三元组抽取的推理代码，见predict_pt.py，其他任务可以根据自己的评价标准进行推理预测。\n### **Lora方法**\nLora方法，即在大型语言模型上对指定参数增加额外的[低秩矩阵](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%BD%8E%E7%A7%A9%E7%9F%A9%E9%98%B5&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkvY7np6nnn6npmLUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Ov4N1t7fgwDtcjeDBqTM__ZGhV693r-2AMpVceEqjSA&zhida_source=entity)，并在模型训练过程中，仅训练而外增加的参数。当“秩值”远小于原始参数维度时，新增的低秩矩阵参数量很小，达到仅训练很小的参数，就能获取较好的结果。 \n![](https://pic1.zhimg.com/v2-e9965e231005d772d5ea5a6d2351fe60_1440w.jpg)\n * Lora论文：[Paper](https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2106.09685)\n * 官方代码：[Github](https://link.zhihu.com/?target=https%3A//github.com/microsoft/LoRA)\n * HuggingFace封装的peft库：[Github](https://link.zhihu.com/?target=https%3A//github.com/huggingface/peft)\n\n\n微调代码，见finetuning_lora.py，核心部分如下：\n\n```\nmodel = ChatGLMForConditionalGeneration.from_pretrained(args.model_dir)\nconfig = LoraConfig(r=args.lora_r,\n lora_alpha[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=lora_alpha&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJsb3JhX2FscGhhIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Ns3tttuwY6f77vbrhvwUQJub2T9-eqliRyXxYb0P3Gw&zhida_source=entity)=32,\n target_modules=[\"query_key_value\"],\n lora_dropout=0.1,\n bias=\"none\",\n task_type=\"CAUSAL_LM\",\n inference_mode=False,\n )\n\nmodel = get_peft_model[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=peft_model&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJwZWZ0X21vZGVsIiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6MjI2MTA1ODczLCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.usI5RIgeF7S8wLJomUWgRCifrtAbNpdBOTWeuYBRaKk&zhida_source=entity)(model, config)\n```\n\n可设置参数包含train_path、model_dir、num_train_epochs、train_batch_size、gradient_accumulation_steps、output_dir、prompt_text、lora_r等，可根据自己的任务配置。\n\n```\nCUDA_VISIBLE_DEVICES=0 deepspeed finetuning_lora.py --num_train_epochs 5 --train_batch_size 2 --lora_r 8\n```\n\n三元组抽取的推理代码，见predict_lora.py，其他任务可以根据自己的评价标准进行推理预测。\n注意：对于结果需要保持一致的任务(即关掉dropout，解码关掉do_sample)，需要保存模型的adapter_config.json文件中，inference_mode参数修改成false，并将模型执行model.eval()操作。 主要原因是chatglm模型代码中，没有采用Conv1D函数。\n## **三元组抽取实验结果**\n * 模型训练时，最大长度为768，Batch大小为2，训练轮数为5，fp16训练，采用DeepSpeed的Zero-1训练；\n * PT为官方的P-Tuning V2[训练方法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%AE%AD%E7%BB%83%E6%96%B9%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLorq3nu4Pmlrnms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.6rN-pl1ivWQHYI94TBQ8UO9tm4iUCsWdKUpWhjKoMp4&zhida_source=entity)，PT-Only-Embedding表示仅对Embedding进行soft-prompt，Freeze仅训练模型后五层参数，Lora采用低秩矩阵方法训练，秩为8；\n * 由于之前训练PT在48G-A40显卡上会出现OOM，因此之前进行PT实验时对模型开启了gradient_checkpointing_enable，使得模型显存占用变小，但训练时长增加。\n * 训练示例：\n\n\n\n```\nprompt_text：你现在是一个信息抽取模型，请你帮我抽取出关系内容为\\\"性能故障\\\", \\\"部件故障\\\", \\\"组成\\\"和 \\\"检测工具\\\"的相关三元组，三元组内部用\\\"_\\\"连接，三元组之间用\\\\n分割。文本：\n输入：故障现象：发动机水温高，风扇始终是低速转动，高速档不工作，开空调尤其如此。\n输出：发动机_部件故障_水温高\\n风扇_部件故障_低速转动\n```\n\n时间换空间，可用很好的解决显卡的资源问题，简单玩玩还可以，如果想要模型达到最优效果或可用快速看到效果，还不如租张A100卡，快速实验，推理阶段再用自己的小破卡。\n下面实验结果均是在租的80G-A100上进行的实验，与Github里用的A40的实验结果会有些差异，主要在训练时长(纯[训练速度](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%AE%AD%E7%BB%83%E9%80%9F%E5%BA%A6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLorq3nu4PpgJ_luqYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.t-AmRGbze4gLsjexQKbxq-v0D80rV6tiq80Ul2u100s&zhida_source=entity)，剔除模型保存的时间)。说实话，真的要训练一个大模型，多个A100是必不可少的，可以减少很多模型并行的操作，效果上也更好把控一些。 \n| 微调方法 | PT-Only-Embedding | PT | Freeze | Lora | \n| --- | --- | --- | --- | --- | \n| 显卡占用 | 37G | 56G | 24G | 39G | \n| 总参数 | 6.259B | 7.211B | 6.255B | 6.259B | \n| 可训练参数占比 | 0.0586% | 13.26% | 16.10% | 0.0586% | \n| 训练耗时 | 20min | 52min | 46min | 25min | \n| 测试结果F1 | 0.0 | 0.6283 | 0.5675 | 0.5359 | \n结果分析：\n * 效果为PT>Freeze>Lora>PT-Only-Embedding;\n * 速度为PT-Only-Embedding>Lora>Freeze>PT;\n * **PT-Only-Embedding效果很不理想，发现在训练时，最后的loss仅能收敛到2.几，而其他机制可以收敛到0.几。分析原因为，输出内容形式与原有语言模型任务相差很大，仅增加额外Embedding参数，不足以改变复杂的[下游任务](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%B8%8B%E6%B8%B8%E4%BB%BB%E5%8A%A1&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkuIvmuLjku7vliqEiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.689UbNINqmQlBd2lHtDgyjruko4c9-orNeyeGV78czQ&zhida_source=entity)**;\n * PT方法占用[显存](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=2&q=%E6%98%BE%E5%AD%98&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmmL7lrZgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MiwiemRfdG9rZW4iOm51bGx9.CtSACxtivtl_MMRqfYJP8ICzvtt1wIiE6XfN-z-FVyo&zhida_source=entity)更大，因为也增加了很多而外参数;\n * 测试耗时，采用float16进行[模型推理](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%A8%A1%E5%9E%8B%E6%8E%A8%E7%90%86&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmqKHlnovmjqjnkIYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.IULSkFKFuuVYRH-uC3v4vqUEDytRGKvdXVpjMRJ0epQ&zhida_source=entity)，由于其他方法均增加了额外参数，因此其他方法的推理耗时会比Freeze方法要高。当然由于是生成模型，所以生成的长度也会影响耗时;\n * 模型在指定任务上微调之后，并没有丧失原有能力，例如生成“帮我写个[快排算法](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%BF%AB%E6%8E%92%E7%AE%97%E6%B3%95&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlv6vmjpLnrpfms5UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.UFSJHLHGt7WiEp4XQrrFoxx2NQOZ0cPUhSxivPzdedQ&zhida_source=entity)”，依然可以生成-快排代码;\n * **由于[大模型微调](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=2&q=%E5%A4%A7%E6%A8%A1%E5%9E%8B%E5%BE%AE%E8%B0%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlpKfmqKHlnovlvq7osIMiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MiwiemRfdG9rZW4iOm51bGx9.sWKd6yIQXkr3qGoYR2V231GCQUP3mCdubzL-ITAfzDU&zhida_source=entity)都采用大量instruction进行模型训练，仅采用单一的指令进行微调时，对原来其他的指令影响不大，因此并没导致原来模型的能力丧失**;\n * 上面测试仅代表个人测试结果。\n\n\n很多同学在微调后出现了灾难性遗忘现象，但我这边并没有出现，对“翻译任务”、“代码任务”、“问答任务”进行测试，采用freeze模型，可以用test_forgetting.py进行测试，具体测试效果如下：\n * 翻译任务 \n\n\n![](https://pic4.zhimg.com/v2-6d4d29a974c10411a915b51fb226e919_1440w.png)\n * 代码任务 \n\n\n![](https://pic4.zhimg.com/v2-bd9a3e20503f83d7f57b5dabf98fef43_1440w.jpg)\n * 问答任务 \n\n\n![](https://pic4.zhimg.com/v2-b0901617f0084a1bffdeeb1961fc05d3_1440w.png)\n后面会把生成任务、分类任务做完，请持续关注Github，会定期更新。（太忙了，会抓紧时间更新，并且官方代码也在持续更新，如遇到代码代码调不通的情况，请及时联系我，我在github也给出了我的代码版本和模型版本）\n### **文本生成**\n * 为了防止大模型的数据泄露，采用一个“万创杯”[中医药天池大数据竞赛](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E4%B8%AD%E5%8C%BB%E8%8D%AF%E5%A4%A9%E6%B1%A0%E5%A4%A7%E6%95%B0%E6%8D%AE%E7%AB%9E%E8%B5%9B&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLkuK3ljLvoja_lpKnmsaDlpKfmlbDmja7nq57otZsiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.VByrU5xRrBIyvUU6GvrupUs9fCE1XWUA1Dp0b71D_Ws&zhida_source=entity)-[中医文献问题生成挑战](https://link.zhihu.com/?target=https%3A//tianchi.aliyun.com/competition/entrance/531826/introduction)，随机抽取20条作为测试集\n * PT为官方的P-Tuning V2训练方法，PT-Only-Embedding表示仅对Embedding进行soft-prompt，Freeze仅训练模型后五层参数，Lora采用低秩矩阵方法训练，秩为8；\n * 训练示例：\n\n\n\n```\nprompt_text：你现在是一个问题生成模型，请根据下面文档生成一个问题，文档：\n输入：紫色红薯是近年从日本[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%97%A5%E6%9C%AC&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLml6XmnKwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.-32xvuaLamIVhtZxcDg4pWwnBjqrFSTn-_x3IXcorik&zhida_source=entity)引进的新品种红薯，中国农业大学农学与生物技术学院副院长刘庆昌[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%88%98%E5%BA%86%E6%98%8C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLliJjluobmmIwiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.ydJi2gCYBnZesDRrNQ_RZgHtG03KbZi7rhBvP-jgtss&zhida_source=entity)指出，紫薯中的花青素[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%8A%B1%E9%9D%92%E7%B4%A0&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLoirHpnZLntKAiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.A9NiU9LOrtSRTRk9QU9KuEKxeXgFk-zPEMaBGceWJxk&zhida_source=entity)具有显著的抗生物[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%8A%97%E7%94%9F%E7%89%A9&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmipfnlJ_niakiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.4qoI7eNOv9Hcu9s09Bthm3fExl2e3FoTa2sVnp8xpmA&zhida_source=entity)氧化作用，在延缓人体衰老方面具有非常好的效果。紫薯中所含赖氨酸[](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E8%B5%96%E6%B0%A8%E9%85%B8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLotZbmsKjphbgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.tJA7ZRuA34EHtSoYnrKL_oeuz00EmmE2wFVqrS4B3Ak&zhida_source=entity)、铜、锰、钾、锌的含量高于一般红薯5-8倍，尤其是抗癌物质碘、硒的含量比其他红薯高出20倍以上，占食物中的第一位。\n输出：紫薯和红薯吃哪个好？\n```\n\n[模型训练](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=4&q=%E6%A8%A1%E5%9E%8B%E8%AE%AD%E7%BB%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmqKHlnovorq3nu4MiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6NCwiemRfdG9rZW4iOm51bGx9.da8PBQ0lNUyi0mwwnoACUdwYNotYbeZk7GhgKpXfZ1k&zhida_source=entity)，以freeze方法为例：\n\n```\nCUDA_VISIBLE_DEVICES=0 nohup deepspeed --master_port 5555 finetuning_freeze.py --train_path \"data/d2q_0.json\" --output_dir \"output_dir_freeze/\" --prompt_text \"你现在是一个问题生成模型，请根据下面文档生成一个问题，文档：\" > log_fz.log 2>&1 &\n```\n\n由于生成模型的内容不能想信息抽取任务一样评价，用现有的BLUE或者[Rouge](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=Rouge&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJSb3VnZSIsInpoaWRhX3NvdXJjZSI6ImVudGl0eSIsImNvbnRlbnRfaWQiOjIyNjEwNTg3MywiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.hUI4f9AWlnY2-5_5tlVgmWzULKBbbr1KSD6slj7UFiU&zhida_source=entity)来评价也是不合适，因此制定了评分规则。 通过多样性和准确性两个角度判断D2Q模型好坏，每个样本总计5分，共20个样本。\n * 多样性： \n * 问题是否高度相似，每重复一个问题扣0.25分；\n * 问题对应答案是否相同，每有一个重复答案或找不到答案，扣0.25分；\n * 准确性： \n * 问题能否从文档中找到答案，每有一个找不到答案，扣0.25分；\n * 问题内容是否流畅，每有一个问题不流畅，扣0.25分；\n * 问题内容是否有害，每有一个有害，扣0.25分；\n\n\n测试数据见d2q_result_data/，测试代码见predict_d2q.py \n| 微调方法 | 原始模型 | PT-Only-Embedding | PT | Freeze | Lora | \n| --- | --- | --- | --- | --- | --- | \n| 分数 | 51.75 | 73.75 | 87.75 | 79.25 | 86.75 | \n## **中文开源大模型 &项目**\n虽然出来很多大模型，但Open的&中文可直接使用的并不多，下面对中文开源大模型、数据集和项目进行一下汇总。 \n![](https://pic4.zhimg.com/v2-3422e1fd8f182d35418d30216a079905_1440w.jpg)\n### **中文开源大模型**\n直接可微调，无需指令[增量训练](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E5%A2%9E%E9%87%8F%E8%AE%AD%E7%BB%83&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLlop7ph4_orq3nu4MiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.aszCUw9y4N-vuDpdPwOapcfp38bMXsfp98OSkhY9-GA&zhida_source=entity)：\n * ChatGLM-6B：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/THUDM/chatglm-6b)\n * ChatYuan-large-v2：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/ClueAI/ChatYuan-large-v2)\n\n\n原始模型多语言or英文，需要中文指令数据集增量训练：\n * BloomZ：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/bigscience/bloomz)\n * LLama：[模型地址](https://link.zhihu.com/?target=https%3A//github.com/facebookresearch/llama)\n * Flan-T5：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/google/flan-t5-xxl)\n * OPT：[模型地址](https://link.zhihu.com/?target=https%3A//huggingface.co/facebook/opt-66b)\n\n\n### **中文开源指令数据**\n下面中文[指令集](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=%E6%8C%87%E4%BB%A4%E9%9B%86&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiLmjIfku6Tpm4YiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.Je9Z6VreHMwaGkowyOWtcGDslowmn7Q6r8OoufT-3fs&zhida_source=entity)，大多数从Alpaca翻译而来，请看下面项目中data目录。目前通过ChatGPT或者GPT4作为廉价标注工为自己的数据进行数据标注一个不错的思路。\n * [1]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/LC1332/Chinese-alpaca-lora)\n * [2]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/hikariming/alpaca_chinese_dataset)\n * [3]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/carbonz0/alpaca-chinese-dataset)\n * [4]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM)\n * [5]：[数据地址](https://link.zhihu.com/?target=https%3A//github.com/LianjiaTech/BELLE)\n * [6]：[数据地址](https://link.zhihu.com/?target=https%3A//huggingface.co/datasets/JosephusCheung/GuanacoDataset)\n\n\n### **开源项目**\n总结下面较火的开源项目：\n * BELLE：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/LianjiaTech/BELLE)\n * ChatGLM：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/THUDM/ChatGLM-6B)\n * Luotuo-Chinese-LLM：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/LC1332/Luotuo-Chinese-LLM)\n * stanford_alpaca：[项目地址](https://link.zhihu.com/?target=https%3A//github.com/tatsu-lab/stanford_alpaca)\n\n\n## **总结**\n目前各大厂的大模型陆陆续续放出，堪称百家争鸣！个人玩家也是全面拥抱，想尽一切办法来训练微调大模型。只愿大家以后可以实现“大模型”自由。愿再无“[model-as-a-service](https://zhida.zhihu.com/search?content_id=226105873&content_type=Article&match_order=1&q=model-as-a-service&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzkxNDEsInEiOiJtb2RlbC1hcy1hLXNlcnZpY2UiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjoyMjYxMDU4NzMsImNvbnRlbnRfdHlwZSI6IkFydGljbGUiLCJtYXRjaF9vcmRlciI6MSwiemRfdG9rZW4iOm51bGx9.6bAH0u0Hx6-tNterCdAl8vv8rlEpNDWKi_1f4shS2oI&zhida_source=entity)”。\n* * *\n**欢迎多多转发，点赞，关注，有问题的朋友也欢迎加我微信** 「logCong」**私聊，交个朋友吧，一起学习，一起进步。**\n**我们的口号是“生命不止，学习不停”。**\n## **往期回顾**\n[刘聪NLP：ChatGPT-所见、所闻、所感](https://zhuanlan.zhihu.com/p/605331104)\n[刘聪NLP：AAAI2023 | 基于统一语义匹配的通用信息抽取框架-USM](https://zhuanlan.zhihu.com/p/598882471)\n[刘聪NLP：CC-Riddle：汉字谜语问答数据集](https://zhuanlan.zhihu.com/p/537552626)\n[刘聪NLP：中文NER数据集整理](https://zhuanlan.zhihu.com/p/529541521)\n[刘聪NLP：ACL2022 | DCSR：一种面向开放域段落检索的句子感知的对比学习方法](https://zhuanlan.zhihu.com/p/527366495)\n[刘聪NLP：ACL2022 | NoisyTune：微调前加入少量噪音可能会有意想不到的效果](https://zhuanlan.zhihu.com/p/523865674)\n[刘聪NLP：ACL2022|DictBERT：通过低频词典增强预训练模型表征方法](https://zhuanlan.zhihu.com/p/523551515)\n[刘聪NLP：ACL2022论文分类汇总-Prompt、句子表征、检索排序&摘要](https://zhuanlan.zhihu.com/p/519943388)\n[刘聪NLP：总结|Prompt在NER场景的应用](https://zhuanlan.zhihu.com/p/518146549)\n[刘聪NLP：PolyLoss：一种将分类损失函数加入泰勒展开式的损失函数](https://zhuanlan.zhihu.com/p/510626670)\n[刘聪NLP：PERT：一种基于乱序语言模型的预训练模型](https://zhuanlan.zhihu.com/p/509647368)\n[刘聪NLP：DiffCSE：结合句子间差异的无监督句子嵌入对比学习方法](https://zhuanlan.zhihu.com/p/507171467)\n[刘聪NLP：SIGIR2022论文筛选](https://zhuanlan.zhihu.com/p/506005620)\n[刘聪NLP：PairSCL：句子对级别的有监督对比学习方法](https://zhuanlan.zhihu.com/p/463949437)\n[刘聪NLP：COLD：中文冒犯性语言检测数据集](https://zhuanlan.zhihu.com/p/463455280)\n[刘聪NLP：SNCSE：一种基于软负例的无监督句向量对比学习方法](https://zhuanlan.zhihu.com/p/463142612)\n[刘聪NLP：CPT模型：一种中文兼顾NLU和NLG的非平衡预训练语言模型](https://zhuanlan.zhihu.com/p/456553852)\n[刘聪NLP：SimCSE论文精读](https://zhuanlan.zhihu.com/p/452761704)\n[刘聪NLP：常用预训练语言模型（PTMs）总结](https://zhuanlan.zhihu.com/p/406512290)\n[刘聪NLP：ACL2021论文之ChineseBERT：融合字形与拼音信息的中文预训练模型](https://zhuanlan.zhihu.com/p/393617564)\n[刘聪NLP：Sentence-Bert论文笔记](https://zhuanlan.zhihu.com/p/113133510)\n[刘聪NLP：MacBERT：MLM as correction BERT](https://zhuanlan.zhihu.com/p/250595837)\n[刘聪NLP：智能扩充机器人的“标准问”库之Query生成](https://zhuanlan.zhihu.com/p/149429784)\n[刘聪NLP：短文本相似度算法研究](https://zhuanlan.zhihu.com/p/111414376)\n编辑于 2023-06-30 12:44・江苏\n[LLM（大型语言模型）](https://www.zhihu.com/topic/26797383)\n[自然语言处理](https://www.zhihu.com/topic/19560026)\n[人工智能](https://www.zhihu.com/topic/19551275)\n​赞同 1731​​167 条评论​4044 ​318 \n​分享\n​申请转载​\n​\n![](https://pic1.zhimg.com/v2-d41c2ceaed8f51999522f903672a521f_l.jpeg)\n未登录用户\n \n\n167 条评论\n默认\n最新\n[![刘聪NLP](https://pica.zhimg.com/v2-53ee42573c43af1ebc16953b5dbd8768_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n[刘聪NLP](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n作者​![](https://picx.zhimg.com/v2-27bfcba90e66db79ce8768ab807e017e_l.png?source=32738c0c)\n更新了文本生成的实验![\\[吃瓜\\]](https://pic4.zhimg.com/v2-74ecc4b114fce67b6b42b7f602c3b1d6.png)\n2023-04-18 · 江苏 · 作者置顶\n​回复​8\n[![徐晓龙](https://picx.zhimg.com/9c3602ca9af7226b93b3337b4bbe5568_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/6785657c61e5032307fce9e7f7c5ad9e)\n[徐晓龙](https://www.zhihu.com/people/6785657c61e5032307fce9e7f7c5ad9e)\n你这个里面没看到没微调前的模型效果，有这块的数据嘛\n2023-04-14 · 安徽\n​回复​8\n[![Szymon](https://picx.zhimg.com/v2-801c29dfbe1dcf8e256dc296a422b6e8_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/4eb0e2b49dd64734145f896f3303b982)\n[Szymon](https://www.zhihu.com/people/4eb0e2b49dd64734145f896f3303b982)\n这里有条挺重要的，专项训练过后是否影响原先的各种性能。\n2023-04-20 · 四川\n​回复​5\n展开其他 4 条回复​\n[![半糖](https://pica.zhimg.com/v2-abed1a8c04700ba7d72b45195223e0ff_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/a55028acdb8c97237a8fba9940a2ae95)\n[半糖](https://www.zhihu.com/people/a55028acdb8c97237a8fba9940a2ae95)\n大佬能做下文本分类的微调吗～最近在做感觉效果很差，不知道哪里出问题了55\n2023-05-17 · 福建\n​回复​5\n展开其他 5 条回复​\n[![揽睿星舟](https://picx.zhimg.com/v2-81ff6fdf5cf4f243b9aacebba9b40d70_l.jpg?source=06d4cd63)](https://www.zhihu.com/org/9bdb83ab64dc4b31656f2c14f8f48eff)\n[揽睿星舟](https://www.zhihu.com/org/9bdb83ab64dc4b31656f2c14f8f48eff)\n妈妈！我上电视了！![\\[爱\\]](https://pic1.zhimg.com/v2-0942128ebfe78f000e84339fbb745611.png)\n2023-04-11 · 上海\n​回复​4\n[![刘聪NLP](https://picx.zhimg.com/v2-53ee42573c43af1ebc16953b5dbd8768_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n[刘聪NLP](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n作者​![](https://picx.zhimg.com/v2-27bfcba90e66db79ce8768ab807e017e_l.png?source=32738c0c)\n![\\[酷\\]](https://pic4.zhimg.com/v2-c96dd18b15beb196b2daba95d26d9b1c.png)![\\[酷\\]](https://pic4.zhimg.com/v2-c96dd18b15beb196b2daba95d26d9b1c.png)\n2023-04-11 · 江苏\n​回复​喜欢\n展开其他 1 条回复​\n[![NyChatAI](https://pica.zhimg.com/v2-abed1a8c04700ba7d72b45195223e0ff_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/4ac9362ebdbb221ab13ebaa9b0180a75)\n[NyChatAI](https://www.zhihu.com/people/4ac9362ebdbb221ab13ebaa9b0180a75)\n为什么指令微调的输入是input+label：输出是input的长度*100；能采用input，输出直接是label吗？\n2023-07-24 · 广东\n​回复​2\n[![Lemon](https://pic1.zhimg.com/300d4671fa49678922abc0e9620b55d3_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/b2cab50cbd39c8138c9a7f0a06f341a8)\n[Lemon](https://www.zhihu.com/people/b2cab50cbd39c8138c9a7f0a06f341a8)\n楼主好，想请教一下，关系抽取任务没有影响模型原来的自然语言理解能力，那生成任务有影响嘛？感觉关系抽取任务可能在之前预训练的时候，用到了类似的数据或者指令，所以没有对原始模型产生太大的影响。如果是一个完全特定领域的生成任务，会有影响嘛？\n2023-05-19 · 北京\n​回复​2\n[![wyl](https://picx.zhimg.com/v2-abed1a8c04700ba7d72b45195223e0ff_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/5cc2745fb4c468616a02e712e520c97d)\n[wyl](https://www.zhihu.com/people/5cc2745fb4c468616a02e712e520c97d)\n当prefix_projection为True时，为P-Tuning-V2方法，在大模型的Embedding和每一层前都加上新的参数；为False时，为P-Tuning方法，仅在大模型的Embedding上新的参数。这个有错误吧？True代表的是使用了重参数吧？\n2023-05-18 · 北京\n​回复​2\n展开其他 4 条回复​\n[![王二叔](https://picx.zhimg.com/v2-2ad5759be3b010d3221681a72d126076_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/e630a433d5d02d31e3182c6507612fea)\n[王二叔](https://www.zhihu.com/people/e630a433d5d02d31e3182c6507612fea)\n感谢分享 \n2023-04-12 · 北京\n​回复​1\n[![刘聪NLP](https://pica.zhimg.com/v2-53ee42573c43af1ebc16953b5dbd8768_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n[刘聪NLP](https://www.zhihu.com/people/4338c2384a026f9c1f430d81771f0e0a)\n作者​![](https://picx.zhimg.com/v2-27bfcba90e66db79ce8768ab807e017e_l.png?source=32738c0c)\n![\\[吃瓜\\]](https://pic4.zhimg.com/v2-74ecc4b114fce67b6b42b7f602c3b1d6.png)\n2023-04-12 · 江苏\n​回复​喜欢\n点击查看全部评论\n![](https://pic1.zhimg.com/v2-d41c2ceaed8f51999522f903672a521f_l.jpeg)\n未登录用户\n \n\n关于作者\n[![刘聪NLP](https://pic1.zhimg.com/v2-53ee42573c43af1ebc16953b5dbd8768_l.jpg?source=32738c0c&needBackground=1)](https://www.zhihu.com/people/LiuCongNLP)\n[刘聪NLP](https://www.zhihu.com/people/LiuCongNLP)\nLLM is all you need！\n[​](https://zhuanlan.zhihu.com/p/96956163)\n云问科技 算法工程师\n[ 回答 **505** ](https://www.zhihu.com/people/LiuCongNLP/answers)[ 文章 **205** ](https://www.zhihu.com/people/LiuCongNLP/posts)[ 关注者 **35,939** ](https://www.zhihu.com/people/LiuCongNLP/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[江歌妈妈发长文尘埃终将落定706 万](https://www.zhihu.com/search?q=%E6%B1%9F%E6%AD%8C%E5%A6%88%E5%A6%88%E5%8F%91%E9%95%BF%E6%96%87%E5%B0%98%E5%9F%83%E7%BB%88%E5%B0%86%E8%90%BD%E5%AE%9A&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[中国男足 0-5 巴勒斯坦564 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[原央视主持人阿丘回应被通报422 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[华为与赛力斯达成新五年合作411 万](https://www.zhihu.com/search?q=%E5%8D%8E%E4%B8%BA%E4%B8%8E%E8%B5%9B%E5%8A%9B%E6%96%AF%E8%BE%BE%E6%88%90%E6%96%B0%E4%BA%94%E5%B9%B4%E5%90%88%E4%BD%9C&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[迪拜航空确认航班发生事故400 万](https://www.zhihu.com/search?q=%E8%BF%AA%E6%8B%9C%E8%88%AA%E7%A9%BA%E7%A1%AE%E8%AE%A4%E8%88%AA%E7%8F%AD%E5%8F%91%E7%94%9F%E4%BA%8B%E6%95%85&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[东航回应网传空姐跪地道歉327 万](https://www.zhihu.com/search?q=%E4%B8%9C%E8%88%AA%E5%9B%9E%E5%BA%94%E7%BD%91%E4%BC%A0%E7%A9%BA%E5%A7%90%E8%B7%AA%E5%9C%B0%E9%81%93%E6%AD%89&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n[王楚钦：不知道为什么就是感觉累320 万](https://www.zhihu.com/search?q=%E7%8E%8B%E6%A5%9A%E9%92%A6%EF%BC%9A%E4%B8%8D%E7%9F%A5%E9%81%93%E4%B8%BA%E4%BB%80%E4%B9%88%E5%B0%B1%E6%98%AF%E6%84%9F%E8%A7%89%E7%B4%AF&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)热\n[江苏高考作文《衬衫的价格为 9 磅 15 便士》爆火268 万](https://www.zhihu.com/search?q=%E6%B1%9F%E8%8B%8F%E9%AB%98%E8%80%83%E4%BD%9C%E6%96%87%E3%80%8A%E8%A1%AC%E8%A1%AB%E7%9A%84%E4%BB%B7%E6%A0%BC%E4%B8%BA+9+%E7%A3%85+15+%E4%BE%BF%E5%A3%AB%E3%80%8B%E7%88%86%E7%81%AB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n[女装网店开始用防拆带了265 万](https://www.zhihu.com/search?q=%E5%A5%B3%E8%A3%85%E7%BD%91%E5%BA%97%E5%BC%80%E5%A7%8B%E7%94%A8%E9%98%B2%E6%8B%86%E5%B8%A6%E4%BA%86&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n[孕妇骑车别车被司机踹翻258 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=6daa0213-369d-41dd-9af2-40c394bc724e)\n### 推荐阅读\n# [llm微调记录 03 微调方法LlamaFactory 的微调方法有3种： Full： \\theta_{new}=\\theta_{old}+\\Delta\\theta Lora: \\theta_{new}=\\theta_{old}+B\\cdot A^T Freeze:冻结层： \\theta_{new}=\\theta_{old} 训练层… 谁知道呢 ](https://zhuanlan.zhihu.com/p/1907910324352456379)[ ![模型应用-LLM微调](https://picx.zhimg.com/v2-168324d306d4bbf18223cd2551afea41_250x0.jpg?source=172ae18b)模型应用-LLM微调 Dev-Z...发表于模型方法 ](https://zhuanlan.zhihu.com/p/676490419)[ ![LLM微调实践和总结](https://pic1.zhimg.com/v2-2b47ecbf7536521771b789216cd207b7_250x0.jpg?source=172ae18b)LLM微调实践和总结 yuwei1991 ](https://zhuanlan.zhihu.com/p/684763523)[ ![大模型学习笔记（四） 更好的LLM：微调、提示词工程与对齐](https://pic1.zhimg.com/v2-834288ef8f1e97fe1da5762249737e32_250x0.jpg?source=172ae18b)大模型学习笔记（四） 更好的LLM：微调、提示词工程与对齐 水滴油穿 ](https://zhuanlan.zhihu.com/p/714988018) \n\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](https://picx.zhimg.com/v2-9e41ea16bdfbe9cf4896617ecad5b4ca.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n", "body_base_path": "attachments/e0eda9767a9f/pages/", "body_markdown_path": "attachments/e0eda9767a9f/pages/archived-body.md", "local_html": "attachments/e0eda9767a9f/pages/archived-body.html", "body_source": "prior_capture" }