minimind-deep-dive
Chinese-language study notes that walk through the MiniMind LLM codebase line by line, then extend each concept into the wider techniques used across modern large language models.
minimind-deep-dive is a set of Chinese-language study notes built around reading the MiniMind project's source code section by section: each note names the exact source file and symbol it discusses, explains what that code does and why, and is meant to be read side by side with the MiniMind repository. Beyond the code walkthrough, each topic is extended outward into the broader technique it belongs to, for example going from RMSNorm to the evolution of normalization methods, or from CISPO to the family of GRPO variants, plus an appendix of chapters covering topics MiniMind does not touch at all, such as quantization, speculative decoding, and the history of model generations.
The main line of ten chapters covers MiniMind's structure, its training stages from pretraining through PPO/GRPO, the underlying training mechanics from logits to a parameter update, a chapter comparing two versions of MiniMind, and a chapter of real experiments run on the author's own server, including training curves and evaluation results, with the observation that reward increasing during RL training does not necessarily mean factual or code correctness improves on held-out evaluation.