Snowball-67B-A2B-10T-Mixed-RLVR2-Async-Step146

Snowball 67B-A2B — 10T Mixed RLVR2 async step 146

This is the Hugging Face export of optimizer checkpoint 146 from the September 11 Snowball bounded-staleness asynchronous mixed-domain RLVR lineage. RLVR1 was followed by RLVR2 without data shuffling. The arm used RLOO-N, maximum staleness 2, TIS, router replay, TITO, the Megatron backend, a 32K context window, batch size 512, and 16 rollouts per prompt on 64 H100s.

Checkpoint 140 was the true RLVR2 training-pass@16 peak at 0.666015625. It was removed by an earlier checkpoint pruning sweep, and the S3 bucket had no recoverable object version. Step 146 is the best complete retained checkpoint; its training pass@16 was 0.63671875, training reward was 0.59610817, and its holdout avg_score/pass@1 was 0.4625/0.37.

This checkpoint is a research artifact. The async lineage includes early local inference-bridge overload, an amended holdout set after LiveCodeBench was removed, and RLVR2 validation that did not improve over the stronger RLVR1 region.

• 10T Mixed RLVR collection
• Animated per-domain RLVR1 rewards
• Public experiment artifacts

We release our tools to the public with the intent that they are used safely, ethically, and responsibly. They are meant to contribute to the public's understanding of science, and aid in its advancement. By using this Marin model, you agree to comply with all applicable law.