{"data":{"slug":"baichuan-inc-baichuan2","name":"Baichuan2","tagline":"A series of large language models developed by Baichuan Intelligent Technology","github_url":"https://github.com/baichuan-inc/Baichuan2","owner":"baichuan-inc","repo":"Baichuan2","owner_avatar_url":"https://avatars.githubusercontent.com/u/136167093?v=4","primary_language":"Python","stars":4084,"forks":294,"topics":["artificial-intelligence","benchmark","ceval","chatgpt","chinese","gpt","gpt-4","huggingface","large-language-models","llama2","mmlu","natural-language-processing"],"archived":false,"github_pushed_at":"2024-11-08T08:31:29+00:00","maintenance_label":"Dormant","stars_delta_30d":-6,"url":"https://www.graphcanon.com/tools/baichuan-inc-baichuan2","markdown_url":"https://www.graphcanon.com/tools/baichuan-inc-baichuan2.md","api_url":"https://www.graphcanon.com/api/graphcanon/tools/baichuan-inc-baichuan2","graph_url":"https://www.graphcanon.com/api/graphcanon/graph?tool=baichuan-inc-baichuan2","description":"A series of large language models developed by Baichuan Intelligent Technology","homepage_url":"https://huggingface.co/baichuan-inc","license":"Apache-2.0","open_issues":281,"watchers":39,"ai_summary":"Baichuan 2 is a new generation of open-source large language models developed by Baichuan Intelligent Technology, trained on high-quality corpus with 2.6 trillion tokens.","readme_excerpt":"
\n

\n Baichuan 2\n

\n
\n\n

\n🤗 Hugging Face • 🤖 ModelScope • 💬 WeChat• 🧩 Modelers\n

\n\n
\n\n
\n\n
\n\n🚀 百小应APP已正式上线！懂搜索、会提问的AI助手，各大应用商店搜索百小应下载，欢迎下载体验 🎉\n\n\n\n\n\n

\n

\n 中文 |\n English\n

\n

\n\n
\n\n# 目录\n\n- [📖 模型介绍](#模型介绍)\n- [📊 Benchmark 结果 🥇🥇🔥🔥](#Benchmark-结果)\n- [⚙️ 推理和部署](#推理和部署)\n- [🛠️ 模型微调](#模型微调)\n- [💾 中间 Checkpoints 🔥🔥](#中间-Checkpoints)\n- [👥 社区与生态](#社区与生态)\n- [📜 声明、协议、引用](#声明协议引用)\n\n# 更新\n[2023.12.29] 🎉🎉🎉 我们发布了 **[Baichuan2-13B-Chat v2](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/tree/v2.0)** 版本。其中：\n- 大幅提升了模型的综合能力，特别是数学和逻辑推理、复杂指令跟随能力。\n\n# 模型介绍\n\n- Baichuan 2 是百川智能推出的**新一代开源大语言模型**，采用 **2.6 万亿** Tokens 的高质量语料训练。\n- Baichuan 2 在多个权威的中文、英文和多语言的通用、领域 benchmark 上取得同尺寸**最佳**的效果。\n- 本次发布包含有 **7B**、**13B** 的 **Base** 和 **Chat** 版本，并提供了 Chat 版本的 **4bits 量化**。\n- 所有版本对学术研究完全开放。同时，开发者通过邮件申请并获得官方商用许可后，即可**免费商用**，请参考[协议](#协议)章节。\n- 欢迎阅读我们的技术报告 [Baichuan 2: Open Large-scale Language Models](https://arxiv.org/abs/2309.10305) 获取更多信息。\n\n本次发布版本和下载链接见下表：\n| | 基座模型 | 对齐模型 | 对齐模型 4bits 量化 |\n|:-------:|:-------:|:-------:|:-----------------:|\n| 7B | 🤗 [Baichuan2-7B-Base](https://huggingface.co/baichuan-inc/Baichuan2-7B-Base) | 🤗 [Baichuan2-7B-Chat](https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat) | 🤗 [Baichuan2-7B-Chat-4bits](https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat-4bits) |\n| 13B | 🤗 [Baichuan2-13B-Base](https://huggingface.co/baichuan-inc/Baichuan2-13B-Base) | 🤗 [Baichuan2-13B-Chat](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat) | 🤗 [Baichuan2-13B-Chat-4bits](https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat-4bits) |\n\n# Benchmark 结果\n\n我们在[通用](#通用领域)、[法律](#法律医疗)、[医疗](#法律医疗)、[数学](#数学代码)、[代码](#数学代码)和[多语言翻译](#多语言翻译)六个领域的中英文和多语言权威数据集上对模型进行了广泛测试。\n\n## 通用领域\n\n在通用领域我们在以下数据集上进行了 5-shot 测试。\n- [C-Eval](https://cevalbenchmark.com/index.html#home) 是一个全面的中文基础模型评测数据集，涵盖了 52 个学科和四个难度的级别。我们使用该数据集的 dev 集作为 few-shot 的来源，在 test 集上进行测试。我们采用了 [Baichuan-7B](https://github.com/baichuan-inc/Baichuan-7B/tree/main) 的评测方案。\n- [MMLU](https://arxiv.org/abs/2009.03300) 是包含 57 个任务的英文评测数据集，涵盖了初等数学、美国历史、计算机科学、法律等，难度覆盖高中水平到专家水平，是目前主流的 LLM 评测数据集。我们采用了[开源](https://github.com/hendrycks/test)的评测方案。\n- [CMMLU](https://github.com/haonan-li/CMMLU) 是一个包含 67 个主题的综合性性中文评估基准，专门用于评估语言模型在中文语境下的知识和推理能力。我们采用了其[官方](https://github.com/haonan-li/CMMLU)的评测方案。\n- [Gaokao](https://github.com/OpenLMLab/GAOKAO-Bench) 是一个以中国高考题作为评测大语言模型能力的数据集，用以评估模型的语言能力和逻辑推理能力。 我们只保留了其中的单项选择题，并进行了随机划分。我们采用了与 C-Eval 类似的评测方案。\n- [AGIEval](https://github.com/microsoft/AGIEval) 旨在评估模型的认知和解决问题相关的任务中的一般能力。 我们只保留了其中的四选一单项选择题，并进行了随机划分。我们采用了与 C-Eval 类似的评测方案。\n- [BBH](https://huggingface.co/datasets/lukaemon/bbh) 是一个挑战性任务 Big-Bench 的子集。Big-Bench 目前包括 204 项任务。任务主题涉及语言学、儿童发展、数学、常识推理、生物学、物理学、社会偏见、软件开发等方面。BBH 是从 204 项 Big-Bench 评测基准任务中大模型表现不好的任务单独拿出来形成的评测基准。\n\n### 7B 模型结果\n\n| | **C-Eval** | **MMLU** | **CMMLU** | **Gaokao** | **AGIEval** | **BBH** |\n|:---------------------:|:----------:|:--------:|:---------:|:----------:|:-----------:|:-------:|\n| | 5-shot | 5-shot | 5-shot | 5-shot | 5-shot | 3-shot |\n| **GPT-4** | 68.40 | 83.93 | 70.33 | 66.15 | 63.27 | 75.12 |\n| **GPT-3.5 Turbo** | 51.10 | 68.54 | 54.06 | 47.07 | 46.13 | 61.59 |\n| **LLaMA-7B** | 27.10 | 35.10 | 26.75 | 27.81","github_created_at":"2023-08-31T08:42:37+00:00","created_at":"2026-07-07T17:35:31.741444+00:00","updated_at":"2026-08-17T18:01:31.04106+00:00","categories":[{"slug":"inference-serving","name":"Inference & Serving","url":"https://www.graphcanon.com/categories/inference-serving","markdown_url":"https://www.graphcanon.com/categories/inference-serving.md","api_url":"https://www.graphcanon.com/api/graphcanon/categories/inference-serving"},{"slug":"llm-frameworks","name":"LLM Frameworks","url":"https://www.graphcanon.com/categories/llm-frameworks","markdown_url":"https://www.graphcanon.com/categories/llm-frameworks.md","api_url":"https://www.graphcanon.com/api/graphcanon/categories/llm-frameworks"}],"tags":[{"slug":"artificial-intelligence","name":"artificial-intelligence"},{"slug":"benchmark","name":"benchmark"},{"slug":"chatgpt","name":"chatgpt"},{"slug":"chinese","name":"chinese"},{"slug":"huggingface","name":"huggingface"},{"slug":"large-language-models","name":"large language models"},{"slug":"natural-language-processing","name":"natural-language-processing"}],"trust":{"provenance":{"is_fork":false,"github_id":685444668,"owner_type":"Organization","methodology":"github_public_v1","parent_repo":null,"near_duplicate_slugs":[]},"computed_at":"2026-08-17T18:01:28.707Z","maintenance":{"label":"Dormant","score":18,"methodology":"github_public_v1","releases_90d":0,"days_since_push":647,"last_release_at":null,"stars_delta_30d":-6,"open_issues_delta_30d":-2},"security_summary":{"status":"ok","scanner":"osv@v1","low_count":0,"high_count":0,"last_scan_at":"2026-07-11T11:05:36.011Z","medium_count":0,"scan_profile":"deps","critical_count":0}},"capability_facts":{"scan":{"source":"repo_scan","observed_at":"2026-08-17T18:01:29.928Z"},"languages":{"value":["python"],"source":"github.language","observed_at":"2026-08-17T18:01:29.928Z"},"license_spdx":{"value":"Apache-2.0","source":"github.license","observed_at":"2026-08-17T18:01:29.928Z"}},"decision_facts":{"hosting":null,"pricing":null,"requirements":null,"constraints":null,"when_to_use":["- Requires robust language processing capabilities, particularly for Chinese, as Baichuan2 has demonstrated superior performance on comprehensive Chinese language assessments such as C-Eval and CMMLU."],"when_not_to_use":["- Needs real-time feedback or interactive deployment in low-resource environments where high computational requirements may be prohibitive, given its large model size and complexity."],"source":"enrich:decision_facts","observed_at":"2026-07-09T22:00:51.367Z"},"constraint_facets":null,"decision_summary":[{"label":"Adopt for","value":"- When to use: Excellent for tasks requiring deep understanding in multiple languages, especially Chinese. Suitable for applications that demand high performance across a wide range of benchmarks and domains.\n\n- When not"}]}}