jingyaogong/minimind

完整覆蓋預訓練、SFT、LoRA、RLHF、MoE、多模態等流程，且核心程式碼用PyTorch原生實作，適合學習LLM內部機制。README提供詳細模型列表、資料集、訓練腳本與第三方框架相容性，文件完整度高。只需單張消費級GPU即可復現，上手成本低。 Covers pretraining, SFT, LoRA, RLHF, MoE, and multimodal pipelines with PyTorch-native implementations, ideal for learning LLM internals. README provides detailed model lists, datasets, training scripts, and third-party compatibility, with high documentation completeness. Runs on a single consumer GPU, lowering onboarding cost.