From collaborative filtering to diffusion models, one repo

What it does

Fun-Rec is an open-source Chinese textbook—self-styled the “Wheat Book”—that teaches recommendation systems from the ground up. It covers the classical cascade pipeline, from collaborative filtering and vector recall to ranking and rerank, then advances into generative territory: LLM-based reasoning, diffusion models, and hardware-aware scaling laws. The repository also includes a complete production-grade system with offline and online components.

The interesting bit

Rather than treating generative recommendation as a speculative research thread, the authors frame it as the next engineering chapter, dedicating full sections to industrial architectures like OneRec, HSTU, and even “thinking” models such as RecZero. It is essentially a community course trying to outrun the standard curriculum.

Key highlights

• Starts with classical recall and ranking methods, then jumps to sequence modeling, multi-task learning, and diversity reranking.
• Devotes half its syllabus to generative paradigms: tokenizer design for recsys, Scaling Law exploration, and end-to-end retrieval generation.
• Surveys concrete industrial proposals including OneSearch, OneSug, and diffusion-based augmentations like DiffuASR.
• Bundles a full production-level build covering system architecture, offline pipelines, online flow, and deployment.
• Maintained by the Datawhale community under a CC BY-NC-SA 4.0 license.

Caveats

• The project is a work in progress; the authors note that content and code change frequently and they are not currently accepting pull requests.
• The core text is in Chinese, so non-Mandarin readers will need to rely on browser translation or the sparse English README.

Verdict

A strong bookmark for Mandarin-speaking developers who want a structured path from traditional recsys engineering to frontier generative methods. Look elsewhere if you need a polished, final-edition reference.