Step38 science SFT: Science-forward
Archival Hugging Face export from the Science-forward science SFT run at optimizer step 13,161. The run started from Snowball Step38. Its GPQA-Diamond result was 55.22%, versus 58.25% for the Step38 starting model.
The experiment issue, public score and configuration archive, and draft reproducibility PR document this checkpoint. The physical input dataset is being uploaded and is currently incomplete. The model is a research artifact, and the raw-science SFT arms did not establish a broad improvement over Step38. Original upstream license terms still apply.