The Chinese LLM tutorial that makes you hand-craft Llama 2

What it does

This is an open-source Chinese textbook-style repository covering the full stack from classical NLP to LLM engineering. It walks through tokenizers, RNNs, Transformers, and BERT/GPT, then advances to fine-tuning Qwen2.5 with LoRA, RLHF, quantization, and serving models with FastAPI and Docker. The material is organized as a six-part syllabus with hands-on projects like text classification and named-entity recognition.

The interesting bit

The unusual angle is its insistence on hand-crafting core architectures—readers literally write Llama 2 and Transformer blocks from scratch rather than importing them—while still pushing through to the unglamorous but necessary engineering bits like Jenkins CI/CD and model quantization. It is essentially trying to turn enthusiasts into engineers who can both derive an attention head and containerize a model service.

Key highlights

• Covers the full arc from Word2Vec and RNNs to MoE, RLHF, and multimodal models like BLIP-2 and LLaVA.
• Explicitly rejects the “package-caller” mindset by making readers hand-write Llama 2 and Transformer core code.
• Practical deployment chapters cover FastAPI, Docker Compose, cloud server setup, and Jenkins pipelines.
• Uses a “question-iteration-refactoring” teaching style, showing how scripts evolve into industrial-grade code.
• Published under a CC BY-NC-SA 4.0 license, so it is explicitly non-commercial and share-alike.

Caveats

• The LLM Safety section is incomplete: two of its chapters are marked as under construction.
• The Extra-chapter is currently an empty placeholder.
• All content is in Chinese, so non-Chinese readers need translation support.

Verdict

Care if you are a Chinese-speaking developer, student, or researcher who wants to close the gap between NLP theory and LLM engineering without skipping the math or the DevOps. Skip it if you are looking for a quick API reference or a production-ready framework to import; this is a curriculum, not a library.