FineEnvs/Qwen3.5-2B-FineProof-RL

Text Generation • 2B • Updated about 4 hours ago • 3

Note Qwen3.5-2B after 500 RL updates on FineProofEnvs: held-out proofs 0.217 → 0.407 (gpt-oss-120b regrade 0.259 → 0.464).

Note FineProofEnvs train: 5,099 olympiad proof tasks (2,899 validated) from QED-Nano's release, as Harbor tasks with a rubric judge.

Note FineProofEnvs test: 128 olympiad proof tasks (64 validated) as Harbor tasks, graded 0–7 against each rubric.

Note Trackio dashboard: the 500-update run, 14 held-out evals with gpt-oss-120b regrades, and every eval attempt as a trace.

Note Every held-out eval attempt (proof, gpt-oss-20b grades, 120b regrade), every training update and rollout of the run, as tables.

Note openai/math: research-level Lean 4 proofs, checked by Comparator in a separate sandbox.

Note Scott Wu Math: make the target from every card once with + − × ÷. Play it, or point an agent at it.