jiaran-king/re-zero---starting-llm-

This repository is a personal knowledge base for studying large language models, built as an Obsidian vault and written in Chinese. It is not a piece of software you install or run. Instead, it is a structured set of notes the author uses to organize what they learn about LLMs: research papers, engineering frameworks, and hands on experiments. The notes are arranged in layers. Raw source material gets logged first, then distilled into an index, then rewritten into standalone concept notes, then linked together into topic maps, and finally connected to project and experiment write ups. Folders reflect this pipeline, with separate directories for the homepage and index, topic maps, concept notes, project notes, experiment logs, source excerpts, note templates, and an inbox for material that has not been sorted yet. A small Python script in the tools folder helps convert Obsidian style links and canvas diagrams into a form that renders properly on GitHub. The topics covered include core transformer architecture (attention, tokenization, positional encoding), pretraining and scaling laws, parameter efficient fine tuning methods like LoRA and QLoRA, alignment techniques such as RLHF, PPO, and DPO, and inference optimization topics like KV caching and Flash Attention. There are also notes on training infrastructure tools such as vLLM and Ray, plus summaries of specific papers and models like Qwen and DeepSeek. A new reader is pointed toward a suggested reading order: start with the overview page, move to the learning map to pick a topic, then drill into that topic's map and its underlying concept notes, tracing back to source material only when needed. The repository includes a maintenance guide describing how new material should be filed so the structure stays organized as it grows. This is best understood as a public study log rather than a tool or library. There is no installed application, API, or command line interface to use. The repository does not specify an open source license, so its contents should not be treated as freely reusable without checking with the author first.