Prototype Policy Explorer
FairTune Evaluation Gates
Explore how utility, safety, and fairness criteria could govern model promotion—without presenting illustrative inputs as measured model improvements.
View GitHubCandidate Report
Illustrative gate logicPromotion Decision
Adjust a candidate report to inspect the prototype's promotion-gate policy.
No model, dataset, Firebase service, or external API is called.
Validation Boundary
A useful workflow prototype, not a fairness result.
FairTune is independently built and its Python source compiles locally. It has no automated tests, has not been production tested, and currently leaves fairness evaluation unimplemented.
The utility evaluator targets a small SQuAD validation subset using exact match and F1.
Validation is currently limited to source compilation and inspection of the included evaluation report.
Repository Evidence
A QLoRA scaffold using Transformers, PEFT, and a small Alpaca subset.
Exact-match and F1 evaluation against a 50-sample SQuAD slice.
Six fixed prompts scored for toxicity, threats, and insults with Detoxify.
Counterfactual parity is documented as a goal but remains a placeholder in code.