Snowball 67B A2B Antidoom RLVR1, step 12
This is the Hugging Face export of the Antidoom RLVR1 policy at optimizer step 12. The run started from the Antidoom FTPO/DAPO step-45 export and trained on broad-domain Nemotron Ultra RLVR tasks using 80 policy and 48 inference H100 GPUs. The 100-prompt holdout reached 34% pass@1 and 0.3705 average verifier score at this step. A later continuation of the same run collapsed, so step 12 was selected before export.
The run record and figures, configuration gist, and W&B history provide provenance. The step-12 weights were strictly loaded from a preserved distributed checkpoint before conversion; the pinned September 21 Datakit SFT model supplied compatible architecture metadata when the old step-45 export was unavailable regionally.
Research artifact. This model has not been properly tested or evaluated and is not necessarily secure. Use at your own risk in production settings.