A local coding model for ordinary programmer work. Embabel is the first domain that can be scored with a compiler and a planner. Spring and Kotlin patterns, taken from real example source, are part of the same training set.
This repository is the model program. The MCP bridge that serves a local model to Cursor, Copilot, and Claude stays in slm-setup. Training does not go into that bridge.
- docs/research-plan.md — research question, phases, evaluation, and the controlled base / retrieval / LoRA experiment.
- docs/milestones.md — the production bar and the milestone order. LoRA waits until retrieval has been measured on the same tasks.
- docs/m0-bars.md — the frozen general-coding floor, Embabel gain, and host budget. A later run does not edit this page after the numbers look bad.
- docs/paper-calendar-2026.md — Nov 2026–Mar 2027 preprint plan (Zenodo primary; TechRxiv contingency).
- docs/citation-workflow.md — formal citation acquisition, triage, and BibTeX rules.
- docs/paper/outline.md — Paper v1 claim freeze and section skeleton.
Pin these before any adapter run. The same pins feed retrieval and verified training examples. Holdout tasks stay out of the training split.
| Source | What to pin |
|---|---|
| Embabel Guide | Official embabel/guide. Read it for training. Do not open a pull request against it. |
| Embabel DICE | Official embabel/dice |
| Embabel documentation | Official agent guide on docs.embabel.com |
| Learning sites | jmjava/embabel-v1-learning (cheat sheet, walkthrough, lessons) |
| Embabel cookbook | Official cookbook on docs.embabel.com |
| Spring patterns | Pattern cards mined from Spring usage in official embabel/guide, official embabel/dice, and the learning demos. Each card cites the source file it came from. |
| Kotlin patterns | Pattern cards mined from Kotlin in official embabel/dice and the learning demos. Each card cites the source file it came from. |
Skip .env files, tokens, and local secrets when ingesting. A training example that only restates a doc page, with no code a compiler can accept or reject, does not count. A Spring or Kotlin pattern card with no source path stays out.
- Do not start with fine-tuning.
- Do not overwrite an experiment directory.
- Do not choose the permanent base model from a single chat session.
- Official Ollama library tags only, until a later decision says otherwise.
- Model weights are not committed. See models/README.md.