Snowball 67B-A2B, R2E-Gym RL, new stack, step 24
RL on the migrated marin stack (harbor + MarinSkyRL, 2026-09-11/12), 1,003-task tt-v2-train curriculum pool, hidden tests, no KL term. Held-out pass@1: .600 same-repo, .595 unseen-repo, .079 never-solved (base .375 / .402 / .036). The selected checkpoint of the 2026-09-12 wave.
Base: laion/snowball-67b-a2b-sft-s3-nemotron-terminal-step1888 (Snowball 67B-A2B, GrugMoe). Training: RLOO-style on-policy RL (sequence-mean loss, staleness 2, groups of 8, batch 64 prompts, lr 5e-7, bf16 stochastic rounding) on R2E-Gym tasks with the terminus-2 terminal agent, 49,152-token prompt / 16,384-token generation. Held-out numbers are pass@1 at 8 attempts on the tt-v2 val441 split (150 same-repo, 115 unseen-repo, 176 held-out), paired against the base model.
Files: HF-layout safetensors export of the policy weights (39 shards), config and tokenizer. Serve with the GrugMoe vLLM fork; the EAGLE-3 draft laion/snowball-64k-eagle3-draft-r2egym gives ~1.5x decode on this family.
- Downloads last month
- 10