HypoDiverse: GRPO
This is the exact merged Hugging Face checkpoint evaluated on HypoDiverse.
Provenance
| Field | Value |
|---|---|
| method | GRPO |
| Base model | Qwen/Qwen3-4B |
| Dataset | viciousa3gis/hypodiverse |
| Pinned dataset revision | d16867cc49836f72ace9e3667164fa6e4ae76eda |
| Evaluation protocol | standard |
This checkpoint is the validity-reward GRPO baseline. Each completion is rewarded for producing a hypothesis that is consistent with the visible evidence; the reward has no explicit set-diversity term.
Exact training and evaluation configurations, per-file model hashes, and the
pinned dataset revision are recorded in release_manifest.json and
provenance/configs/.