This repository is a Mac-first tutorial project for training a tiny 20-30M parameter decoder-only LLM from scratch on local hardware. It keeps deep-research-report.md as the research foundation and layers a practical docs-plus-code implementation on top.
The main path uses TinyStories for the first successful run. An optional second track uses Hacker News for curated tech-domain adaptation after the base model is already working.
Start with docs/01-overview-and-prereqs.md.