What it does

Happy-LLM is a free, Chinese-language course in notebook form that walks you through building and training a large language model from the ground up. It starts with NLP fundamentals and Transformer theory, then moves to hands-on implementation: you’ll build a working LLaMA2 architecture, train a tokenizer, run pretraining, and apply SFT, LoRA, and QLoRA fine-tuning. The final chapters cover RAG and Agent applications.

The interesting bit

Most LLM courses hand you a framework and say “call .fit().” This one makes you write the PyTorch layers first, then introduces Hugging Face Transformers only after you’ve felt the pain. There’s even a published 215M-parameter checkpoint you can download and compare against your own attempt.

Key highlights

• Seven completed chapters from NLP basics through RAG and Agents; an Extra Chapter collects community blog posts and PRs
• Explicit per-chapter Python environments to avoid dependency hell
• Pretrained and SFT-tuned 215M models hosted on ModelScope with a live demo space
• PDF and PPT teaching materials available, watermarked specifically to deter resellers
• Backed by Datawhale, a Chinese AI open-source community, with CCF-recognized AI普惠 courses offering free compute

Caveats

• Entirely in Chinese; English speakers need to rely on browser translation or look elsewhere
• The “偏好对齐” (preference alignment) section in Chapter 6 is marked WIP
• Extra Chapter blog posts are community-contributed and quality varies; maintainers curate what gets merged into the main text

Verdict

Ideal for Mandarin-speaking developers who want to understand LLMs by getting their hands dirty with matrix multiplications before touching high-level frameworks. If you already ship production models via API calls and don’t care about the internals, this will feel like homework.