🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
[update] readme
jingyaogongcacf1d4cd081
diff --git a/README.md b/README.mdindex 39b4184..6c54591 100644--- a/README.md+++ b/README.md@@ -34,7 +34,7 @@* 此开源项目旨在完全从 0 开始,仅用 3 块钱成本与 2 小时训练时间,即可训练出规模约为 64M 的超小语言模型 MiniMind。* MiniMind 系列极其轻量,主线最小版本体积约为 GPT-3 的 $\frac{1}{2700}$,力求让普通个人 GPU 也能快速完成训练与复现。* 项目同时开源了大模型的极简结构与完整训练链路,覆盖 MoE、数据清洗、预训练(Pretrain)、监督微调(SFT)、LoRA、RLHF(DPO)、RLAIF(PPO / GRPO / CISPO)、Tool Use、Agentic RL、自适应思考与模型蒸馏等全过程代码。-* MiniMind 同时拓展了视觉多模态版本 [MiniMind-V](https://github.com/jingyaogong/minimind-v)。+* MiniMind 同时拓展了视觉多模态版本 [MiniMind-V](https://github.com/jingyaogong/minimind-v)、扩散语言模型(MiniMind-dLM)、线性模型(MiniMind-Linear),详见 [Discussion](https://github.com/jingyaogong/minimind/discussions)。* 项目所有核心算法代码均从 0 使用 PyTorch 原生实现,不依赖第三方库提供的高层抽象接口。* 这不仅是一个大语言模型全阶段开源复现项目,也是一套面向 LLM 入门与实践的教程。* 希望此项目能为更多人提供一个可复现、可理解、可扩展的起点,一起感受创造的乐趣,并推动更广泛 AI 社区的进步。@@ -94,6 +94,7 @@- 支持在 C-Eval、C-MMLU、OpenBookQA 等第三方测评集上进行评测,并支持通过 YaRN 实现 RoPE 长文本外推。- 提供兼容 OpenAI API 协议的极简服务端,便于接入 FastGPT、Open-WebUI 等第三方 Chat UI,并支持 `reasoning_content`、`tool_calls`、`open_thinking`。- 提供基于 Streamlit 的极简聊天 WebUI,支持思考展示、工具选择与多轮 Tool Call。+- 包含实验性拓展:离散扩散语言模型([dLM](https://github.com/jingyaogong/minimind/discussions/618))与线性注意力模型([Linear Attention](https://github.com/jingyaogong/minimind/discussions/704)),均可基于主线 AR 模型进行续训。#### 🎉 已发布模型列表diff --git a/README_en.md b/README_en.mdindex e8fffd8..cdc7c01 100644--- a/README_en.md+++ b/README_en.md@@ -34,7 +34,7 @@* This open-source project aims to train an ultra-small language model MiniMind with approximately 64M parameters entirely from scratch, using only 3 CNY in cost and 2 hours of training time.* The MiniMind series is extremely lightweight, with the smallest version on the main branch being approximately $\frac{1}{2700}$ the size of GPT-3, striving to enable even ordinary personal GPUs to quickly complete training and reproduction.* The project also open-sources the minimalist structure and complete training pipeline of large models, covering the entire process code for MoE, data cleaning, Pretraining, Supervised Fine-Tuning (SFT), LoRA, RLHF (DPO), RLAIF (PPO / GRPO / CISPO), Tool Use, Agentic RL, Adaptive Thinking, and Model Distillation.-* MiniMind has also been extended to a visual multimodal version [MiniMind-V](https://github.com/jingyaogong/minimind-v).+* MiniMind has also been extended to a visual multimodal version [MiniMind-V](https://github.com/jingyaogong/minimind-v), a diffusion language model (MiniMind-dLM), and a linear attention model (MiniMind-Linear), See [Discussion](https://github.com/jingyaogong/minimind/discussions) for details.* All core algorithm code in the project is implemented from scratch using native PyTorch, without relying on high-level abstract interfaces provided by third-party libraries.* This is not only a full-stage open-source reproduction project for large language models, but also a tutorial oriented towards LLM introduction and practice.* We hope this project can provide a reproducible, understandable, and extensible starting point for more people, to share the joy of creation together and promote the progress of the broader AI community.@@ -94,6 +94,7 @@ Meanwhile, third-party large model frameworks and tool libraries, such as `trans- Supports evaluation on third-party benchmark suites such as C-Eval, C-MMLU, OpenBookQA, etc., and supports RoPE long context extrapolation through YaRN.- Provides a minimalist server compatible with the OpenAI API protocol, convenient for integrating with third-party Chat UIs such as FastGPT, Open-WebUI, etc., and supports `reasoning_content`, `tool_calls`, `open_thinking`.- Provides a minimalist chat WebUI based on Streamlit, supporting thinking display, tool selection, and multi-turn Tool Call.+- Includes experimental extensions: diffusion language model ([dLM](https://github.com/jingyaogong/minimind/discussions/618)) and linear attention model ([Linear Attention](https://github.com/jingyaogong/minimind/discussions/704)), both can be continued-trained from the main AR model.#### 🎉 Released Model List