// the find
datawhalechina/self-llm
《开源大模型食用指南》针对中国宝宝量身打造的基于Linux环境快速微调（全参数/Lora）、部署国内外开源大模型（LLM）/多模态大模型（MLLM）教程
A Chinese-language tutorial collection for deploying and fine-tuning 50+ open-source LLMs on Linux, targeting students and researchers who can't or won't use API services. Each model gets its own folder with environment setup, FastAPI serving, WebDemo, LangChain integration, and LoRA fine-tuning notebooks. Think of it as a field guide for running models locally, written by and for the Chinese ML community.
Coverage is genuinely broad — Qwen, GLM, DeepSeek, LLaMA, Gemma, and more, all with consistent tutorial structure rather than a random grab-bag. The AMD and Ascend NPU sections are rare; most LLM tutorials assume NVIDIA exclusively, so this fills a real gap for anyone on domestic Chinese hardware. The example projects (Chat-嬛嬛 fine-tuning a character persona, AMChat for advanced math) show complete end-to-end workflows rather than just hello-world inference. Community momentum is real: 60+ contributors, regular commits, and structured issue templates keep the content fresher than most tutorial repos of this type.
The notebook-heavy format is a maintenance liability — diffs are unreadable, code is hard to extract, and broken imports don't surface until someone runs the cell. Tutorial quality varies wildly by contributor: some models have five detailed guides while others (Gemma4) are skeleton folders with a `.keep` file. There's no CI or automated validation, so pinned dependency versions silently rot as model APIs and HuggingFace Hub layouts shift. The copy-paste structure across models means a shared bug (like a tokenizer call pattern that changed) has to be fixed in 50 places, and realistically it won't be.