Dr Doom science SFT, step 5,989
This is the final Hugging Face export of the two-epoch October 2026 science SFT continuation. It started from Dr Doom and trained on 3,477,625 synthetic science conversations. The model is preserved as a negative experimental result: GPQA-Diamond fell from 53.87% to 28.62%, and policy-regraded OlympiadBench fell from 16.21% to 0.67%.
The experiment issue summarizes the findings and limitations. The public report archive contains the paired evaluation records, W&B history, audit, and configurations. Draft Marin PR #9718 preserves the training and export code. The synthetic training dataset remains private pending source and derivative license review.
This checkpoint is an archival research artifact. The evaluation regressions make it unsuitable as a recommended replacement for its starting model.