Part of Pi-of-AI — the "Raspberry Pi of AI." This is the Rules-Baker experiment: a small model with a project's house-style coding rules baked into its weights, so it obeys them with an empty system prompt (no rule-bloat in the context window).

smoke_housestyle.gguf is a pipeline smoke-test, not a finished model. It was LoRA-trained on a tiny mock dataset purely to prove the train → merge → GGUF → run-in-browser loop end to end. It overfits the mock template and does not yet follow arbitrary tasks. The real, task-following bake comes from Stage 3+ teacher data (see the repo roadmap). Published here for transparency and to demonstrate the closing-the-loop test — treat it as a demo, not a tool.

The teacher model sees your coding rules and writes compliant code; the stored student training prompt has the rules stripped out. The student learns to obey silently — so at inference time your IDE context stays clean and fast.

Open rules_baker/web/index.html from the repo, then drag-drop this .gguf in. Same tiny prompt, rules obeyed — that's the closing-the-loop test.

MIT. See the project repo.