中文 | English

ChatPiXiu

Logo

ChatPiXiu: Eat every ChatGPT - Output your own chatbot.

GitHub Contributors Issues Issues GitHub pull requests GitHub stars
开源ChatGPT / 全面 / 轻量 / 易用
文章解读 · 视频解读

> **代码开源,大家放心使用,欢迎贡献! 注意:模型的license取决于模型提供方** - [💥最新讯息](#最新讯息) - [💫OpenNLP计划](#OpenNLP计划) - [💫OpenChat-PiXiu](#ChatPiXiu项目) - [🌟开源ChatGPT调研](#开源ChatGPT调研) - [⛏️使用步骤](#使用步骤) - [📄运行示例](#运行示例) - [📄结果展示](#结果展示) - [🛠️常见报错](#常见报错) - [💐参考资料&致谢](#参考资料&致谢) - [🌟赞助我们](#赞助我们) - [🌈Starchart](#Starchart) - [🏆Contributors](#Contributors) ## 最新讯息 - 2023/04/14:ChatPiXiu项目正式启动: - 开源ChatGPT平替调研汇总 - 2023/05/15:60+开源项目,20+基础模型 - 2023/08/21:由羡鱼的个人项目catqaq迁移至OpenLLMAI ## OpenNLP计划 我们是谁? 我们是**羡鱼智能**【xianyu.ai】,主要成员是一群来自老和山下、西湖边上的咸鱼们,塘主叫作羡鱼,想在LLMs时代做点有意义的事!我们的口号是:**做OpenNLP和OpenX!希望在CloseAI卷死我们之前退出江湖!** 也许有一天,等到GPT-X发布的时候,有人会说NLP不存在了,但是我们想证明有人曾经来过、热爱过!在以ChatGPT/GPT4为代表的LLMs时代,在被CloseAI卷死之前,我们发起了OpenNLP计划,宗旨是OpenNLP for everyone! - 【P0】[OpenTextClassification](https://github.com/catqaq/OpenTextClassification):打造一流的文本分类项目,已开源 - 综述:done - 开源项目:done - papers解读:doing - 炼丹术:done - 【P0】OpenSE:句嵌入,自然语言处理的核心问题之一,doing - 【P0】[ChatPiXiu](https://github.com/catqaq/ChatPiXiu):ChatGPT开源平替及领域适应,doing - 【P1】OpenLLMs:大语言模型,doing - 【P2】OpenTextTagger:文本标注,分词、NER、词性标注等 - OpenX:任重而道远 ## ChatPiXiu项目 ChatPiXiu项目为OpenNLP计划的第2个正式的开源项目,旨在Open ChatGPT for everyone!在以ChatGPT/GPT4为代表的LLMs时代,在被OpenAI卷死之前,做一点有意义的事情!未来有一天,等到GPT-X发布的时候,或许有人会说NLP不存在了,但是我们想证明有人曾来过! ### 1.开发计划 本项目的开发宗旨,打造全面且实用的ChatGPT模型库和文档库。**Eat every ChatGPT - Output your own chatbot!** 目前我们正在启动V1版本的开发,整体的开发计划如下,主要包括了文档和代码两类任务,数据的部分我们暂时将其分散到了各个子任务中。 **V1版本:资料调研+通用最小实现+领域/任务适配** #### 1.1 文档分支 文档分支主要负责项目文档的建设,包括通用技术文档和项目相关文档。 **dev_for_docs**:文档分支,主要负责资料调研(算力有限,有调查才有训练权): 1. 【P0】开源ChatGPT调研:持续更新,doing 2. 【P0】训练技术调研:持续更新,doing 3. 【P0】数据调研:doing 4. 【P1】部署技术调研:TODO 5. 【P2】基础模型调研:目前以LLaMA和GLM为主,doing 6. 【P3】技术解读/教程:doing #### 1.2 代码分支 代码分支,负责具体的开发工作,包括数据处理、算法开发、算法评测等,分成通用最小实现和领域/任务适应两种,具体的: **dev_for_chatmini**:通用最小实现分支,尽可能支持不同的基础模型和训练方式,提供可比较的实现。 1. 【P0】ChatGPT最小复现:完整的RLHF复现SFT-RM-PPO,doing 2. 【P0】适配不同的基座模型 3. 适配不同的PEFT算法 4. 【P2】探索新的训练方式 5. 【P3】探索知识迁移:比如蒸馏 **dev_for_chatzhihu**:知乎及问答领域适配,主要想解决一些知乎使用过程中的痛点,比如问题冗余、回答太多等等。 1. 【P0】收集知乎数据收集及处理 1. SFT数据 2. RLHF数据:答案打分 3. 摘要数据:答案/观点汇总、摘要 2. 【P0】基于知乎数据做SFT 3. 【P1】基于知乎数据做RLHF 4. 【P2】输出知乎LoRA 5. 【P3】和知乎热榜聊天的demo **dev_for_chatzhangsan**:法律领域适配,张三犯了什么罪? 1. 【P0】法律领域数据收集及处理 2. 法律条文解释 3. 【P1】罪名判定:张三犯了什么罪? 更多领域,敬请期待! ChatPiXiu-Eat every ChatGPT - Output your own chatbot! ### 2.加入我们 OpenNLP计划的其他内容尚在筹备中,暂时只开源了本项目和[OpenTextClassification](https://github.com/catqaq/OpenTextClassification)项目。欢迎大家积极参与ChatPiXiu的建设和讨论,一起变得更强! 加入方式: - **项目建设**:可以在前面列出的开发计划中选择自己感兴趣的部分进行开发,建议优先选择高优先级的任务。包括资料调研和算法开发等工作。 - OpenLLM技术交流群:知识在讨论中发展,QQ群:740679327 - 技术分享和讨论:输出倒逼输入,欢迎投稿,稿件会同步到本项目的docs目录和知乎专栏OpenNLP. 同时也欢迎大家积极的参与本项目的讨论https://github.com/catqaq/ChatPiXiu/discussions。 ## 开源ChatGPT调研 ### 1.开源ChatGTP平替 注:开源类ChatGPT/LLM汇总,持续更新中,欢迎贡献! 现已超过60+! | 项目 | 基础模型 | lang | 机构 | 数据集 | license | 介绍 | 备注 | | ------------------------------------------------------------ | --------------------------------------------- | ----------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | | [LLaMA](https://github.com/facebookresearch/llama) | LLaMA | Multi | meta | CCNet [67%], C4 [15%], GitHub [4.5%], Wikipedia [4.5%], Books [4.5%], ArXiv [2.5%], Stack Exchange[2%]. | [Apache-2.0 license](https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE) | 可能是目前开源ChatGPT用的最多的基础模型 | 支持多语言,但以英文为主 | | [stanford_alpaca](https://github.com/tatsu-lab/stanford_alpaca) Alpaca | LLaMA | eng | stanford | [alpaca_data](https://github.com/tatsu-lab/stanford_alpaca/blob/main/alpaca_data.json) | [Apache-2.0 license](https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE) | 指令调优的 LLaMA 模型: An Instruction-following LLaMA Model. 让 OpenAI 的 text-davinci-003 模型以 self-instruct 方式生成 52K 指令样本,SFT | FT模型语言以数据为准 | | [ChatLLaMA](https://github.com/nebuly-ai/nebullvm/tree/main/apps/accelerate/chatllama) | LLaMA | | Nebuly+AI | - | [license](https://github.com/nebuly-ai/nebullvm/blob/main/apps/accelerate/chatllama/LICENSE) | 数据集创建、使用 RLHF 进行高效训练以及推理优化。 | | | [Chinese-LLaMA-Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) | LLaMA | mutli | [ymcui](https://github.com/ymcui) | - | [Apache-2.0 license](https://github.com/ymcui/Chinese-LLaMA-Alpaca/blob/main/LICENSE.md) | Chinese LLaMA & Alpaca LLMs; 中文词表扩充 | | | [alpaca-lora](https://github.com/tloen/alpaca-lora) | LLaMA | | stanford | [LLaMA-GPT4 dataset](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM) | [Apache-2.0 license](https://github.com/tloen/alpaca-lora/blob/main/LICENSE) | [LoRA](https://zhuanlan.zhihu.com/p/620327907) | | | Chinese-alpaca-lora Luotuo-Chinese-LLM | LLaMA | | - | | | LoRA | | | [ChatGLM](https://github.com/THUDM/ChatGLM-6B) | GLM | cn/eng | 清华 | 1T 标识符的中英双语数据 | [Apache-2.0 license](https://github.com/THUDM/ChatGLM-6B/blob/main/LICENSE) | 监督微调、反馈自助、人类反馈强化学习 | [PROJECT.md](https://github.com/THUDM/ChatGLM-6B/blob/main/PROJECT.md) | | [FastChat](https://github.com/lm-sys/FastChat) Vicuna | LLaMA | eng | UC Berkeley, CMU, Stanford, UCSD and MBZUAI | ShareGPT, 70k问答指令数据 | [Apache-2.0 license](https://github.com/lm-sys/FastChat/blob/main/LICENSE) | SFT,使用GPT-4作为评判标准,结果显示Vicuna-13B在超过90%的情况下实现了与ChatGPT和Bard相匹敌的能力。 | | | Chinese-Vicuna | LLaMA | cn | - | - | [Apache-2.0 license](https://github.com/Facico/Chinese-Vicuna/blob/master/LICENSE) | LoRA | | | [EasyLM](https://github.com/young-geng/EasyLM) Koala考拉 | LLaMA multi | eng | UC伯克利 | ChatGPT数据和开源数据(Open Instruction Generalist (OIG)、斯坦福 Alpaca 模型使用的数据集、Anthropic HH、OpenAI WebGPT、OpenAI Summarization) | [Apache-2.0 license](https://github.com/Facico/Chinese-Vicuna/blob/master/LICENSE) | SFT/13B/500k条数据 | | | [ColossalChat](https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat) | LLaMA | | ColossalAI | [InstructionWild](https://github.com/XueFuzhao/InstructionWild):104K bilingual datasets | [LICENSE](https://github.com/hpcaitech/ColossalAI/blob/main/applications/Chat/LICENSE) | SFT-RM-RLHF | | | [ChatRWKV](https://github.com/BlinkDL/ChatRWKV) | RWKV | | [BlinkDL](https://github.com/BlinkDL) | - | [Apache-2.0 license](https://github.com/BlinkDL/ChatRWKV/blob/main/LICENSE) | ChatRWKV is like ChatGPT but powered by RWKV (100% RNN) language model, and open source. | | | [ChatYuan](https://github.com/clue-ai/ChatYuan) | T5 | eng/cn | 元语智能 | PromptClue | [LICENSE](https://github.com/clue-ai/ChatYuan/blob/main/LICENSE) | 基于PromptClue进行了监督微调 | | | [OpenChatKit](https://github.com/togethercomputer/OpenChatKit) | GPT-NoX-20B | | Together+LAION+Ontocord.ai | OIG-43M | [Apache-2.0 license](https://github.com/togethercomputer/OpenChatKit/blob/main/LICENSE) | 60亿参数的审核模型,对不合适或者是有害的信息进行过滤 | | | BELLE | Bloom LLama | cn | [LianjiaTech](https://github.com/LianjiaTech) | 10M-ChatGPT生成的数据 | [Apache-2.0 license](https://github.com/LianjiaTech/BELLE/blob/main/LICENSE) | SFT | | | PaLM-rlhf-pytorch | PaLM | | [lucidrains](https://github.com/lucidrains) | - | [MIT license](https://github.com/lucidrains/PaLM-rlhf-pytorch/blob/main/LICENSE) | RLHF | PaLM太大了 | | [dolly](https://github.com/databrickslabs/dolly) | v1:GPT-J-6B v2:pythia | eng | Databricks | The Pile+databricks-dolly-15k | [MIT license](https://github.com/lucidrains/PaLM-rlhf-pytorch/blob/main/LICENSE) | 参考Alpaca; dolly-v2-12b based on pythia-12b | | | LMFlow | LLaMA | | [OptimalScale](https://github.com/OptimalScale) | | | An Extensible Toolkit for Finetuning and Inference of Large Foundation Models. Large Model for All. LLaMA-7B,一张3090耗时 5 个小时 | | | GPTrillion | - | | - | - | - | 1.5万亿,多模态 | | | [open_flamingo](https://github.com/mlfoundations/open_flamingo) | LLaMA CLIP | | LAION | [Multimodal C4](https://github.com/allenai/mmc4) | [MIT license](https://github.com/mlfoundations/open_flamingo/blob/main/LICENSE) | | | | [baize-chatbot](https://github.com/project-baize/baize-chatbot) | LLaMA | eng | [project-baize](https://github.com/project-baize) | 100k dialogs generated by letting ChatGPT chat with itself. | [GPL-3.0 license](https://github.com/project-baize/baize-chatbot/blob/main/LICENSE) | LoRA | | | [ChatPiXiu](https://github.com/catqaq/ChatPiXiu) | multi | | 羡鱼智能 | - | - | LoRA | 筹备阶段 | | [stackllama](https://huggingface.co/blog/stackllama) | LLaMA | | Hugging Face | - | - | 用RLHF训练LLaMA的实践指南 | | | [Lit-LLaMA](https://github.com/Lightning-AI/lit-llama) | LLaMA | multi | lightening-ai | - | [Apache-2.0 license](https://github.com/Lightning-AI/lit-llama/blob/main/LICENSE) | 重写重训LLaMA,绕开license: Implementation of the LLaMA language model based on nanoGPT. | 可商用版LLaMA | | [OPT](https://arxiv.org/abs/2205.01068) | OPT | eng | meta | - | [MIT license](https://github.com/facebookresearch/metaseq/blob/main/LICENSE) | 当年对标GPT3的模型 | | | [Cerebras-GPT](https://huggingface.co/cerebras/Cerebras-GPT-13B) | Cerebras-GPT | eng | Cerebras | The Pile | [Apache-2.0 license](https://github.com/Cerebras/modelzoo/blob/main/LICENSE) | GPT-3 style; 最小 1.11 亿,最大 130 亿,共 7 个模型 | | | BLOOM | BLOOM | multi | [bigscience](https://huggingface.co/bigscience) | Total seen tokens: 366B | 代码:[Apache-2.0 license ](https://github.com/huggingface/transformers-bloom-inference/blob/main/LICENSE)模型:RAIL License v1.0 | 176B;46 种自然语言(包括中文)和 13 种编程语言 | | | GPT-J | GPT-3 | multi | EleutherAI | The Pile | [apache-2.0](https://huggingface.co/models?license=license:apache-2.0) | based on GPT-3; | | | GPT-2 | | | | | | | | | [RWKV](https://github.com/BlinkDL/RWKV-LM) | [RWKV-LM](https://github.com/BlinkDL/RWKV-LM) | cn/eng | [BlinkDL](https://github.com/BlinkDL) | | | | 纯RNN | | [鹏程・盘古 α](https://www.oschina.net/p/pangu-alpha) | | cn | 鹏城 | 2TB | [Apache License 2.0](https://openi.pcl.ac.cn/PCL-Platform.Intelligence/PanGu-Alpha/src/branch/master/LICENSE) | | | | [鹏程・盘古对话](https://www.oschina.net/p/pangu-dialog) | | cn | 鹏城 | | | | | | [悟道](https://www.oschina.net/p/wudao-model) | | cn/eng | BAAI(智源) | | | 多模态; 1.75 万亿参数; 图文:CogView、BriVL;文本:GLM、CPM、Transformer-XL、EVA、Lawformer;生物:ProtTrans | | | [MOSS](https://www.oschina.net/p/moss) | MOSS | cn/eng | [OpenLMLab](https://github.com/OpenLMLab) | 700B tokens | 代码Apache 2.0,数据CC BY-NC 4.0,模型权重GNU AGPL 3.0 | 支持中英双语和多种插件; 基座moss-moon-003-base | | | [伶荔 (Linly)](https://www.oschina.net/p/linly) | LLaMA | cn | [CVI-SZU](https://github.com/CVI-SZU) | | Apache Licence 2.0 | 33B 的 Linly-Chinese-LLAMA 是目前最大的中文 LLaMA 模型 | | | [华驼 (HuaTuo)](https://www.oschina.net/p/huatuo-llama) | LLaMA | cn/eng | [SCIR-HI](https://github.com/SCIR-HI) | | [Apache-2.0 license](https://github.com/SCIR-HI/Huatuo-Llama-Med-Chinese/blob/main/LICENSE) | | 医学 | | [BBT-2](https://www.oschina.net/p/bbt-2) | | cn/eng | | | | 120 亿参数的通用大语言模型 | | | [CodeGeeX](https://www.oschina.net/p/codegeex) | - | code | 鹏城 | | | 130 亿参数的多编程语言代码生成预训练模型 | code | | [RedPajama](https://www.oschina.net/p/redpajama) | gpt-neox | eng | Together、Ontocord.ai、ETH DS3Lab、斯坦福大学 CRFM、Hazy Research 和 MILA 魁北克 AI 研究所 | 800B/1T | Apache-2.0 | 开源地全面对齐LLaMA的训练数据集 | | | [OpenAssistant](https://www.oschina.net/p/open-assistant) | | eng | [LAION-AI](https://github.com/LAION-AI) | | [Apache-2.0 license](https://github.com/LAION-AI/Open-Assistant/blob/main/LICENSE) | OpenAssistant is a chat-based assistant that understands tasks, can interact with third-party systems, and retrieve information dynamically to do so. | | | [StableLM](https://www.oschina.net/p/stablelm) | pythia | eng | [Stability-AI](https://github.com/Stability-AI) | | [Apache-2.0 license](https://github.com/Stability-AI/StableLM/blob/main/LICENSE) | Stability AI Language Models; max len 4096 | | | [StarCoder](https://www.oschina.net/p/starcoder) | | code | [bigcode-project](https://github.com/bigcode-project) | | [Apache-2.0 license](https://github.com/bigcode-project/starcoder/blob/main/LICENSE) | | code | | [SantaCoder](https://www.oschina.net/p/santacoder) | | code | [bigcode](https://huggingface.co/bigcode) | The Stack(v1.1) | the BigCode OpenRAIL-M v1 license | 轻量级 AI 编程模型,1.1B | code | | [MLC LLM](https://www.oschina.net/p/mlc-llm) | | - | [mlc-ai](https://github.com/mlc-ai) | - | [Apache-2.0 license](https://github.com/mlc-ai/mlc-llm/blob/main/LICENSE) | 本地大语言模型部署;Enable everyone to develop, optimize and deploy AI models natively on everyone's devices. | | | [Web LLM](https://www.oschina.net/p/web-llm) | | | [mlc-ai](https://github.com/mlc-ai) | | [Apache-2.0 license](https://github.com/mlc-ai/web-llm/blob/main/LICENSE) | Bringing large-language models and chat to web browsers. | | | [WizardLM](https://www.oschina.net/p/wizardlm) | LLaMA | eng | [nlpxucan](https://github.com/nlpxucan) | | | Evol-Instruct | | | [YaLM 100B](https://www.oschina.net/p/yalm-100b) | | eng/russian | [yandex](https://github.com/yandex) | | [Apache-2.0 license](https://github.com/yandex/YaLM-100B/blob/main/LICENSE) | | | | [OpenLLaMA](https://www.oschina.net/p/openllama) | | multi | [s-JoL](https://github.com/s-JoL) | | [MIT license](https://github.com/s-JoL/Open-Llama/blob/main/LICENSE) | LLaMA 开源复现版 | | | BiLLa | LLaMA | cn/eng | [Neutralzz](https://github.com/Neutralzz) | | | A Bilingual LLaMA with Enhanced Reasoning Ability | | | pandallm | LLaMA | cn/eng | [dandelionsllm](https://github.com/dandelionsllm) | | Apache-2.0 license | | | | pandalm | | | [WeOpenML](https://github.com/WeOpenML) | | [Apache-2.0 license](https://github.com/WeOpenML/PandaLM/blob/main/LICENSE) | PandaLM:可重现和自动化的语言模型评估 | | | [gpt4all](https://github.com/nomic-ai/gpt4all) | | | [nomic-ai](https://github.com/nomic-ai) | | [MIT license](https://github.com/nomic-ai/gpt4all/blob/main/LICENSE.txt) | gpt4all: an ecosystem of open-source chatbots trained on a massive collections of clean assistant data including code, stories and dialogue | | | [stable-vicuna](https://huggingface.co/CarperAI/stable-vicuna-13b-delta) | | eng | [CarperAI](https://huggingface.co/CarperAI) | | [CC-BY-NC-SA-4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) | StableVicuna-13B is a Vicuna-13B v0 model fine-tuned | | | [MPT](https://huggingface.co/mosaicml/mpt-7b) | MPT | eng | [mosaicml](https://huggingface.co/mosaicml) | 1T tokens of English text and code | Apache-2.0 | ALiBi保证了良好的长度外推性 | | | ImageBind | | 多模态 | meta | | | One embedding space to bind them all. | | | [Phoenix](https://github.com/FreedomIntelligence/LLMZoo) | | multi | CUHK | | | 7B/BLOOMZ + 微调 | | | [ChatPLUG](https://github.com/X-PLUG/ChatPLUG) | | | Alibaba | | | Encoder-Decoder/3.7B | | | [BLOOMZ](https://github.com/bigscience-workshop/xmtf) | | multi | BigScience | | | BLOOM + 多任务微调 | | | [CPM-Ant+](https://github.com/OpenBMB/CPM-Live/tree/cpm-ant-plus/cpm-live) | | cn/eng | OpenBMB | | | 10B/Decoder-only(UniLM) | | | [PaLM](https://ai.googleblog.com/2022/04/pathways-language-model-palm-scaling-to.html) | | multi | Google | | - | pathways/540B | 未开源 | | [PaLM 2](https://blog.google/technology/ai/google-palm-2-ai-large-language-model/) | | multi | Google | | - | pathways; 改进的多语言、推理和编码能力 | 未开源 | | | | | | | | | | ### 2.基础模型 注:基础LLM汇总,持续更新中,欢迎贡献! 现已超过15+!个人的工作和研究兴趣会更关注基础模型相关技术及其应用! | model | Architecture/task | lang | tokenizer | vocab | PE | max len | size | org | data | license | intro | notes | | ------------------------------------------------------------ | --------------------------- | ------ | --------- | ---------------------------------------- | -------------------- | ------------------------------ | --------------------------------------- | --------------------------------------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | | [LLaMA](https://huggingface.co/decapoda-research/llama-7b-hf) | decoder/LM | multi | BBPE | 32K | RoPE | 2048 | 7B/13B/33B/65B | meta | 1.4万亿 tokens | [GPL-3.0 license](https://github.com/facebookresearch/llama/blob/main/LICENSE) | Meta 大语言模型: The model comes in different sizes: 7B, 13B, 33B and 65B parameters. | | | GLM | mix/自回归式填空,prefix LM | cn/eng | multi | multi | RoPE | 2048 for 130B 1024 for ChatGLM | 6B/10B/130B | 智谱 | 中英文token各2000亿 | [Apache-2.0 license](https://github.com/THUDM/ChatGLM-6B/blob/main/LICENSE) | ChatGLM-6B: 1T data | | | GPT1 | decoder/LM | eng | BPE | 40478 | learned | 512 | | OpenAI | BooksCorpus/5 G | [MIT License](https://github.com/openai/finetune-transformer-lm/blob/master/LICENSE) | GPT系列的起源 | 分词部分有特殊处理,详见tokenization_openai.py: - lowercases all inputs, - uses `SpaCy` tokenizer and `ftfy` for pre-BPE tokenization if they are installed, fallback to BERT's `BasicTokenizer` if not. | | [GPT2](https://huggingface.co/gpt2-xl) | decoder/LM | multi | BBPE | 50257 | learned | 1024 | 124M/355M/774M/1.5B | OpenAI | WebText/ 40G | [Modified MIT License](https://github.com/openai/gpt-2/blob/master/LICENSE) | 主打zero-shot | GPT系列开源的最后一舞 | | GPT3 | decoder/LM | multi | BBPE | 50257/davinci | learned | 2048 | 175B | OpenAI | 570G | | sparse attention;主打few-shot/in-context learning | 关上了GPT系列的大门 | | GPT3.5 | decoder/LM | multi | BBPE | 100256 | ~learned | 4096 | | OpenAI | - | - | InstructGPT等一系列模型 | 未开源 | | GPT4 | decoder/LM | multi | BBPE | 100256 | ~learned | 32768 | | OpenAI | - | - | 多模态 | 未开源 | | [BLOOM](https://huggingface.co/bigscience/bloom) | decoder/LM | multi | BBPE | 250,680 | ALiBi | 2048 | 176B | bigscience | [Data Cards](https://huggingface.co/spaces/bigscience/BigScienceCorpus) | RAIL License v1.0 | Modified from Megatron-LM GPT2; StableEmbedding | 包括Training logs | | PaLM | | | | | RoPE | | | | | | | | | Chinchilla | | | | | transformer-XL style | | | | | | | | | OPT | | eng | | | learned | | | | | | | | | [GPT-J](https://huggingface.co/EleutherAI/gpt-j-6b) | decoder/LM | multi | BBPE | 50257/50400† (same tokenizer as GPT-2/3) | RoPE | 2048 | 6B | EleutherAI | The Pile | [apache-2.0](https://huggingface.co/models?license=license:apache-2.0) | based on GPT-3; | | | [Lit-LLaMA](https://github.com/Lightning-AI/lit-llama) | | | | | | | | | | | Implementation of the LLaMA language model based on nanoGPT. | | | [Cerebras-GPT](https://huggingface.co/cerebras/Cerebras-GPT-13B) | decoder | eng | BPE | 50257 | Learned | 2048 | | [Cerebras Systems](https://www.cerebras.net/) | The Pile | Apache 2.0 | The family includes 111M, 256M, 590M, 1.3B, 2.7B, 6.7B, and 13B models. All models in the Cerebras-GPT family have been trained in accordance with Chinchilla scaling laws (20 tokens per model parameter) which is compute-optimal. | | | [RWKV](https://github.com/BlinkDL/RWKV-LM) | RNN | eng/cn | | | None(pure RNN) | 1024/4096/8192 | [multi](https://huggingface.co/BlinkDL) | | the Pile | | 结合了 RNN 和 Transformer 的语言模型,适合长文本,运行速度较快,拟合性能较好,占用显存较少,训练用时较少。RWKV 整体结构依然采用 Transformer Block 的思路,相较于原始 Transformer Block 的结构,RWKV 将 self-attention 替换为 Position Encoding 和 TimeMix,将 FFN 替换为 ChannelMix。其余部分与 Transfomer 一致。 | | | CoLT5 | | | | | T5 bias style | | | | | | | | | MOSS | decoder/LM | cn/eng | | | | 2048 | 16B | [OpenLMLab](https://github.com/OpenLMLab) | 700B tokens | 代码Apache 2.0,数据CC BY-NC 4.0,模型权重GNU AGPL 3.0 | 支持中英双语和多种插件 | | | [MPT](https://huggingface.co/mosaicml/mpt-7b) | decoder/LM | eng | BBPE | 50432 | ALiBi | 2048/65k/84k | 7B | [mosaicml](https://huggingface.co/mosaicml) | 1T tokens+各种FT数据 | Apache-2.0 | Although the model was trained with a sequence length of 2048, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. The model vocabulary size of 50432 was set to be a multiple of 128 (as in MEGATRON-LM) | | | [mt5](https://huggingface.co/google/mt5-xxl) | encoder-decoder | multi | | | | | 1.2B/3.7B/13B | Google | mC4 | | | | | [Wenzhong2.0-GPT2-3.5B-chinese](https://huggingface.co/IDEA-CCNL/Wenzhong2.0-GPT2-3.5B-chinese) | decoder/LM | cn | | | | | 3.5B | IDEA-CCNL | | | | | | CPM-Generate | | cn | | | | | 2.6B | TsinghuaAI | 100GB Chinese training data | | | | | [bloom-zh](https://huggingface.co/Langboat/bloom-1b4-zh) | decoder/LM | cn | BBPE | 46145 | ALiBi | 2048 | 1.4B/2.5B/6.4B | Langboat | - | | 词表裁剪,保留中文: 250880 to 46145 | | | [GPT-2B](https://huggingface.co/nvidia/GPT-2B-001) | decoder/LM | eng | BBPE | | RoPE | 4096 | | HuggingFace+Nvidia | 1.1T tokens | | | | | | | | | | | | | | | | | | ### 3.数据 | dataset | type | 机构 | 大小 | license | 介绍 | 备注 | | --------------------------- | ----------- | ----------------------------------- | ---- | ------- | ---- | ---- | | alpaca_data | Instruction | stanford | | | | | | alpaca_chinese_dataset | \- | hikariming | | | | | | Multilingual Instruction | | Guanaco | | | | | | alpaca_chinese_dataset | | carbonz0 | | | | | | 0.5M+1M chinese instruction | | LianjiaTech | | | | | | shareGPT | | [lm-sys](https://github.com/lm-sys) | | | | | | | | | | | | | ### 4.产品 注:为了文档的完整性,将工业界的ChatGPT也进行了汇总,只做介绍不做比较,以免争议! | model | org | intro | notes | | ------------------------------------------------- | --------- | ----- | ----- | | [ChatGPT-GPT-3.5-turbo](https://chat.openai.com/) | OpenAI | | | | [ChatGPT-GPT-4](https://chat.openai.com/) | OpenAI | | | | Claude | Anthropic | | | | 文心一言 | 百度 | | | | 星火人知大模型 | 讯飞 | | | | ChatGLM | 清华/智谱 | | | | MiniMax | MiniMax | | | | 通义千问 | 阿里 | | | | Bard | Google | | | | | | | | ### 5.训练&部署 #### 5.1 训练 | 框架 | type | 机构 | 兼容性 | license | 介绍 | 备注 | | ------------------------------------------------------------ | ------- | ----------------------------------------- | ---------------------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ---- | | [ColossalAI](https://github.com/hpcaitech/ColossalAI) | general | [hpcaitech](https://github.com/hpcaitech) | 高 | [Apache-2.0 license](https://github.com/hpcaitech/ColossalAI/blob/main/LICENSE) | Colossal-AI: Making large AI models cheaper, faster, and more accessible
支持ChatGPT完整复现 | | | [RLHF](https://github.com/sunzeyeah/RLHF) | RL | [sunzeyeah](https://github.com/sunzeyeah) | 基于transformers库实现 | \- | Implementation of Chinese ChatGPT.
SFT、Reward Model和RLHF | | | trlx | RL | [CarperAI](https://github.com/CarperAI) | 强大的transformer 强化学习库 | [MIT license](https://github.com/CarperAI/trlx/blob/main/LICENSE) | A repo for distributed training of language models with Reinforcement Learning via Human Feedback (RLHF)
不支持自定义预训练模型。 | | | [trl](https://github.com/lvwerra/trl) | RL | Hugging Face | 基于transformers | [Apache-2.0 license](https://github.com/lvwerra/trl/blob/main/LICENSE) | 只要是基于ransformers 库开发的预训练库,均可适配,强烈推荐 | | | [DeepSpeed-Chat](https://github.com/microsoft/DeepSpeedExamples/tree/master/applications/DeepSpeed-Chat) | general | [microsoft](https://github.com/microsoft) | 基于DeepSpeed | [Apache-2.0 license](https://github.com/microsoft/DeepSpeedExamples/blob/master/LICENSE) | **训练速度**大幅提升 | | | [nanoGPT](https://github.com/karpathy/nanoGPT) | GPT | [karpathy](https://github.com/karpathy) | | [MIT license](https://github.com/karpathy/nanoGPT/blob/master/LICENSE) | The simplest, fastest repository for training/finetuning medium-sized GPTs. | | | | | | | | | | #### 5.2 部署 ## 使用步骤:TODO 1.克隆本项目 `git clone https://github.com/catqaq/ChatPiXiu.git` 2.准备数据 3.运行示例 ## 结果展示 ## 常见报错 ## 参考资料&致谢 【OpenLLM 011】ChatPiXiu项目-可能是全网最全的ChatGPT复现调研:54+开源ChatGPT平替项目,15+基础模型,8+ ChatGPT产品! - 羡鱼智能的文章 - 知乎 https://zhuanlan.zhihu.com/p/629364056 开源ChatGPT替代模型项目整理https://zhuanlan.zhihu.com/p/618790279 平替chatGPT的开源方案 https://zhuanlan.zhihu.com/p/618926239ChatGPT/GPT4 开源“平替”汇总https://zhuanlan.zhihu.com/p/621324917 完整版 ChatGPT 克隆方案,开源了!https://zhuanlan.zhihu.com/p/617996976 ColossalChat:完整RLHF平替ChatGPT的开源方案 https://zhuanlan.zhihu.com/p/618048558ChatGPT 开源平替来了,开箱即用!前OpenAI团队打造,GitHub刚发布就揽获800+星https://zhuanlan.zhihu.com/p/613556853 LoRA:大模型的低秩适配-最近大火的lora到底是什么东西?为啥stable diffusion和开源ChatGPT复现都在用?https://zhuanlan.zhihu.com/p/620327907? 成本不到100美元!UC伯克利再开源类ChatGPT模型「考拉」:数据量大没有用,高质量才是王道https://zhuanlan.zhihu.com/p/621078208 ChatGPT平替方案汇总https://zhuanlan.zhihu.com/p/618839784 微软宣布开源 Deep Speed Chat,可将训练速度提升 15 倍以上,哪些信息值得关注?https://www.zhihu.com/question/595311294 总结当下可用的大模型LLMshttps://zhuanlan.zhihu.com/p/611403556 可能是最全的开源 LLM (大语言模型)整理 https://my.oschina.net/oscpyaqxylk/blog/8727824 [资源整理]2023-05-11比较全的LLMs的资源整理 - 迷途小书僮的文章 - 知乎 https://zhuanlan.zhihu.com/p/628637821 中文开源1B以上大模型汇总 - nghuyong的文章 - 知乎 https://zhuanlan.zhihu.com/p/613239726 https://github.com/CLUEbenchmark/SuperCLUE https://github.com/THUDM/ChatGLM-6B 一文汇总开源大语言模型,人人都可以拥有自己的ChatGPT - 无忌的文章 - 知乎 https://zhuanlan.zhihu.com/p/622370602 新发布的一些开源商用模型 - LokLok的文章 - 知乎 https://zhuanlan.zhihu.com/p/627785493 最新发布!截止目前最强大的最高支持65k输入的开源可商用AI大模型:MPT-7B! - 数据学习的文章 - 知乎 https://zhuanlan.zhihu.com/p/627420365 复旦团队大模型 MOSS 开源了,有哪些技术亮点值得关注? - 孙天祥的回答 - 知乎 https://www.zhihu.com/question/596908242/answer/2994534005 暴击专家模型!Meta最新多模态大模型ImageBind已开源 - 新智元的文章 - 知乎 https://zhuanlan.zhihu.com/p/628370318 整理开源可用的中文大模型LLMs - 罗胤的文章 - 知乎 https://zhuanlan.zhihu.com/p/616812772 大模型热点论文:谷歌推出 PaLM 2、Meta 开源 ImageBind - MegEngine Bot的文章 - 知乎 https://zhuanlan.zhihu.com/p/628941792 如何评价Google最新发布的PaLM2,效果反超GPT4? - 一堆废纸的回答 - 知乎 https://www.zhihu.com/question/600311066/answer/3022625910 两大可商用开源大模型同时发布!性能不输LLaMA,羊驼家族名字都不够用了 - 量子位的文章 - 知乎 https://zhuanlan.zhihu.com/p/627454901 ## 赞助我们 我们是谁? 我们是羡鱼智能【xianyu.ai】,主要成员是一群来自老和山下、西湖边上的咸鱼们,塘主叫作羡鱼,想在LLMs做点有意义的事!我们的口号是:做OpenNLP和OpenX!希望在OpenAI卷死我们之前退出江湖! ChatPiXiu项目为羡鱼智能【xianyu.ai】发起的OpenNLP计划的第2个正式的开源项目,旨在Open ChatGPT for everyone!在以ChatGPT/GPT4为代表的LLMs时代,在被OpenAI卷死之前,做一点有意义的事情!未来有一天,等到GPT-X发布的时候,或许有人会说NLP不存在了,但是我们想证明有人曾来过! 本项目第一版由本羡鱼利用业务时间(熬夜)独立完成,受限于精力和算力,拖延至今,好在顺利完成了。如果大家觉得本项目对你的NLP学习/研究/工作有所帮助的话,求一个免费的star! 富哥富姐们可以考虑赞助一下!尤其是算力,**租卡的费用已经让本不富裕的鱼塘快要无鱼可摸了**! image-20230324010955205 ## Starchart [![Star History Chart](https://api.star-history.com/svg?repos=catqaq/ChatPiXiu&type=Date)](https://star-history.com/#catqaq/ChatPiXiu&Date) ## Contributors