Snowball-67B-A2B-Math-RL-E17d-Step8-Repaired

Snowball-67B-A2B-Math-RL-E17d-Step8-Repaired

This is the exact checkpoint used for a reported row in the Snowball 67B-A2B math-RL experiment.

Router state: SFT router-bias repaired control. Closely named older laion/rl-snowball-* repositories may contain raw mutable-router exports that collapse under inference. Do not substitute them for this artifact. Models from this campaign are research checkpoints, not production releases; their utility is limited unless router-bias repair or frozen-router integrity is preserved.

• Arm: E17d repaired control
• Step: 8
• Held-out AIME24 / MATH-500 / OlympiadBench: 3.33 / 37.80 / 5.67
• Selection: Only repaired E17d export; final step 20 remains an unrepaired gap
• Source artifact: s3://marin-us-east-02a/marin/exports/snowball-bias-repaired/rl-snowball-e17d-rno2a-rlvrmath-lossfree-sr-20260825-014634/global_step_8/policy/
• Experiment issue: https://github.com/marin-community/marin/issues/7786
• Evidence archive: https://huggingface.co/datasets/penfever/snowball-67b-a2b-math-rl-artifacts

The scores and evaluation caveats are recorded in MATH_EVALS.md in the evidence archive. Preserve config.json, the tokenizer files, and all shards named by model.safetensors.index.json together.