Required files
8 config, dataset, train log, evals, report, artifact, decisionFactory Run Schema v1
A specialist factory needs proof folders, not vibes. This schema defines the local run directory, required metrics, and decision vocabulary that every public artifact should eventually satisfy.
Headline Numbers
Decisions
6 ship, reject, retry-data, retry-training, retry-eval, parkFirst-class outputs
5 data, training, eval, package, reportCompetitive Context
| System | Metric | Score | Size / Class | Comparable? | Readout |
|---|---|---|---|---|---|
| TinyGPT factory schema | required public run files | 8 | repo-local contract | Direct | Defines the minimum evidence bundle each future public model artifact must carry. |
| Ad hoc model card only | before/after reproducibility | weak | single document | Directional | Useful for release notes, but insufficient for a factory claim without eval JSON, decision, and blockers. |
Direct rows share this artifact's eval setup. Directional rows are useful market context but should not be read as leaderboard claims.
Run folder contract
| File | Purpose | Public relevance |
|---|---|---|
| config.json | Target, base, method, thresholds | Explains what was attempted |
| dataset.json | Sources, rows, filtering, heldout | Provenance |
| eval-baseline.json | Frozen baseline result | Before number |
| eval-candidate.json | Candidate result | After number |
| decision.json | Ship/reject/retry call | Honest release status |
Release Blockers
Needs a canonical rendered example
The schema is real, but the website should show one complete run folder as the public example.
Unblock: Promote the SQL routed result into a small report-only rendered artifact.