Snowball 67B-A2B — 10T Mixed RLVR2 sync step 116
This is the Hugging Face export of optimizer checkpoint 116 from the September 11 Snowball synchronous mixed-domain RLVR lineage. RLVR1 was followed by RLVR2 without data shuffling. The arm used RLOO-N, the Megatron backend, a 32K context window, batch size 512, and 16 rollouts per prompt on 64 H100s.
Checkpoint 118 was the true RLVR2 training-pass@16 peak at 0.63671875. It was removed by an earlier checkpoint pruning sweep, and the S3 bucket had no recoverable object version. Step 116 is the best complete retained checkpoint; its training pass@16 was 0.6171875, training reward was 0.67631429, and its 100-row holdout avg_score/pass@1 was 0.5275/0.44.
This checkpoint is a research artifact. RLVR2 validation did not improve over the stronger RLVR1 region.