Skip to content

[GitHub Trending] jingyaogong/minimind

6.5 relevance
Score Breakdown
technical depth
7
novelty
6
actionability
8
community
7
strategic
4
personal
5

Scored daily by a customisable AI persona to surface the most relevant engineering leadership news.

Training small LLMs from scratch is educational but less directly relevant to senior engineer's focus on agent orchestration and cloud.

AI/ML github.com
🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours. - jingyaogong/minimind
Summary

MiniMind 是一个完全从零开始、仅用 3 元 GPU 租用成本和 2 小时训练时间即可训练出 64M 参数超小语言模型的开源项目,主线结构对齐 Qwen3/Qwen3-MoE 生态。项目覆盖 MoE、数据清洗、预训练、SFT、LoRA、RLHF(DPO)、RLAIF(PPO/GRPO/CISPO)、Tool Use、Agentic RL 和模型蒸馏等完整训练链路,所有核心算法均用 PyTorch 原生实现,不依赖 transformers/trl/peft 等高层封装。最新版本还拓展了视觉模态 MiniMind-V、多模态 Omni 模型 MiniMind-O、扩散语言模型和线性注意力模型,并支持 llama.cpp/vllm/ollama 等推理引擎。