Snowball 67B-A2B — 5.7T Mixed RLVR step 38

Research artifact. This model has not been properly tested or evaluated and is not necessarily secure. Use at your own risk in production settings.

This is the Hugging Face export of optimizer checkpoint 38 from the 5.7T Agentic-start Snowball mixed-domain RLVR1 arm. The arm used synchronous RLOO-N, the Megatron backend, a 32K context window, batch size 512, and 16 rollouts per prompt on 64 H100s.

The checkpoint was selected when export began because its 100-row holdout pass@1 was 0.54. Step 42 later tied that value before the arm was stopped at step 46, so this is not the most recent tied checkpoint. The final arm metrics at step 46 were training reward/pass@16 0.6791/0.6875 and holdout avg_score/pass@1 0.5606/0.5000.

This checkpoint is intended as a research artifact. The campaign exposed output-contract and reward-starvation problems in several weak domains; the model should not be interpreted as uniformly improved across the blend.

• 5.7T Mixed RLVR collection
• Animated per-domain RLVR1 rewards
• Public experiment artifacts

Chat-template update (2026-10-13)

Only chat_template.jinja was replaced with the file from Grug-67B-A2B-Antidoom-RLVR1-Step12-2026.10.02 at revision 50fe2f488aee7730bbccc721024b35229075602e. The tokenizer files and weights were not changed. The previous repository revision remains available by commit hash. This template changes tool declarations/history as well as rendering stored reasoning_content in a think block; it is not merely a display change.

The selection was based on a three-arm Shared100-only diagnostic, all using the same Shared100 weights and benchmark configurations:

The first-call tool probe passed 5/5 in every arm. The historical unmodified Shared100 BFCL score was 59/123 (48.0%); it was not a simultaneous matched control. The one-repeat AIME24 figures have only 30 trials and should not be compared as precise estimates against a ten-repeat score. Both non-selected arms completed all 123 native BFCL trials with zero trial errors, but their Marin roots failed afterward during FineStore S3 aggregation; their BFCL figures are from the native Harbor job aggregates, not canonical Marin records.

The pinned ablation code defines the three arms. Durable BFCL evidence:

• Template-only canonical record: s3://marin-us-east-02a/marin/evals/20261013-102819-shared100-chat-template-template_only-bfclparity-pi-1d7d/record.json
• Template plus tokenizer native job: s3://marin-us-east-02a/marin/evals/20261013-102819-shared100-chat-template-template_and_tokenizer-bfclparity-pi-cdc6/results/harbor_jobs/harbor_bfcl_parity_4eedc64be8a1/result.json
• Reasoning-only native job: s3://marin-us-east-02a/marin/evals/20261013-102641-shared100-chat-template-reasoning_only-bfclparity-pi-7f82/results/harbor_jobs/harbor_bfcl_parity_cbe66a59c30f/result.json

Scope: Only Shared100 was measured in this ablation. Applying its template to Step92 and Step38 is an extrapolation, not evidence that either checkpoint retained benchmark performance after this update. Previous scores and serving provenance refer to the pre-update revisions.

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.