open-athena/Marin-Snowball-GRPO-Results

Marin Snowball GRPO results Two 20-update teacher-repair GRPO pilots completed successfully but did not improve observed held-out identity full-success counts. Scores below count successful conversations; arrows denote baseline → update 20. The control scored identity 1/20 → 0/20 and preservation 19/20 → 20/20. Zero-advantage filtering scored identity 1/20 → 0/20 and preservation 18/20 → 19/20. Both selected the base model; there is no trained model export. Control report… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/Marin-Snowball-GRPO-Results.

<iframe src="https://silytics.si/embed/item/19f694c2-0812-42de-ad60-96f7310eb8bc" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 24 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.