Datawhale Open-Source AI Tutorials (Pumpkin Book, llm-universe, self-llm, happy-llm): Review

Who’s writing this

I’m Lin Zhou, a full-stack developer in China who moved into AI application work and now mentors newcomers on my team. When one of them asks “where do I start learning LLMs for free,” my standing answer includes Datawhale — not out of politeness. During my own career switch, these were the Chinese-language repos that actually carried me through. I’m bilingual, so let me be direct about who this resource is and isn’t for.

What it is

Datawhale is one of China’s most active open-source AI learning communities, and its GitHub matrix is the backbone of zero-cost AI self-study in Chinese. Four projects, each with a clear job. The Pumpkin Book derives roughly 250 of the hardest formulas in Zhou Zhihua’s Machine Learning (the “Watermelon Book”), with errata. llm-universe teaches LLM application development: unified wrappers for GPT, Wenxin, Spark, and GLM APIs, prompt engineering, embeddings and vector databases, RAG with LangChain, Streamlit deployment. self-llm covers local deployment and LoRA/full fine-tuning of open models — Qwen, GLM, InternLM, LLaMA. happy-llm walks from NLP basics to hand-building LLaMA2, the full pretrain/SFT/LoRA pipeline, and agentic RL. All free and open source. Maintenance is healthy as of August 2026: self-llm around 31.8k stars, happy-llm 33.1k, Pumpkin Book 26.0k, llm-universe 13.8k, with commits landing in July and August 2026.

Everything is written in Simplified Chinese — that’s the project’s reason to exist, and the hard boundary for non-readers.

Pick by goal

The Pumpkin Book needs undergraduate math and the Watermelon Book as a prerequisite; it’s a companion, not a standalone text. llm-universe asks only basic Python and no GPU; part one is finished and takes a few hours per the official line. self-llm assumes basic Linux and rentable GPU compute (AutoDL is the common choice, at your own cost). happy-llm assumes Python plus deep learning basics across 8 chapters — 4 theory, 4 practice — with tens of hours of code reproduction. Beyond the tutorials, the community runs free group-study check-in cohorts, including AI summer-camp-style events as of August 2026.

Where it falls short

llm-universe’s parts two and three (advanced RAG, open-source app walkthroughs) have been “in progress” since 2023 — effectively abandoned — and part one’s stack dates quickly, as the open issues reflect. self-llm carries 164 open issues, mostly environment and version traps, and with open-source tutorials nobody is paid to pull you out of a hole. Group-study signup hides behind WeChat official accounts and group chats with no public schedule, and the GitHub team-learning repo went stale in January 2023 — newcomers struggle to find the door. And none of it comes with any certificate.

Alternatives

• Hung-yi Lee’s NTU lectures — listen to Lee for principles, then build with Datawhale; they’re natural complements.
• Hugging Face courses — also free and open, more international in ecosystem, but English; Datawhale is the better Chinese-language experience.
• Baidu AI Studio — pick it for free compute and official labs; pick Datawhale for the PyTorch/open-model ecosystem.

Bottom line

The most systematic, genuinely free, still-maintained open-source AI tutorial collection in Chinese — the deductions are llm-universe’s abandoned chapters and the opaque event signups. If you can self-direct and you read Chinese, this is the ceiling of value, because the price is zero. If you don’t read Chinese, Hugging Face’s courses are your equivalent.