FineEnvs/fineproof-rl-runs

FineProof RL runs

Everything the FineProof RL run produced, in three tables.

• The run: fp-rl2-2b, Qwen3.5-2B trained with RL from the base model (no SFT) on FineProofEnvs for 500 updates.
• The model: published as FineEnvs/Qwen3.5-2B-FineProof-RL.
• Held-out mean reward: 0.217 → 0.407 with the gpt-oss-20b judge, and 0.259 → 0.464 with a gpt-oss-120b regrade.

The same run as charts and traces: the Trackio dashboard. The recipe and the write-up: RESULTS.md.

eval_attempts (default)

One row per proof written in a held-out eval: 64 validated FineProofs-RL-test problems × 4 attempts, for each of the 14 evaluated checkpoints. Step 0 is the base model.

Column
run, model, step	which run and checkpoint
task, attempt	the problem (a FineProofEnvs test task) and the attempt, 0–3
proof	the model's whole reply, as the judge read it
reward	the eval reward: mean of judge_scores / 7
judge_scores, judge_assessments, judge_models	each gpt-oss-20b grade (0–7), its written assessment and the provider that gave it
regrade_120b_score, regrade_120b_assessment	gpt-oss-120b (high effort) regrade, at steps 0, 40, 100, 250, 450 and 500. The assessment keeps its last 1,500 characters
completion_tokens, finish_reason, truncated	the generation; a proof cut off at 32,768 tokens has finish_reason = length
proof_chars, trial	the proof's length, and the Harbor trial it came from

from datasets import load_dataset

attempts = load_dataset("FineEnvs/fineproof-rl-runs", "eval_attempts", split="train")
final = attempts.filter(lambda r: r["run"] == "fp-rl2-2b" and r["step"] == 500)

train_updates

One row per training update: the reward and its spread, KL, entropy, gradient norm and loss from TRL. Also the rollout and judge statistics over the rollouts that finished since the last update: graded share, truncation, tokens, judge disagreement and cost. Plus timings and throughput.

train_rollouts

One row per training rollout (128 per update):

• reward;
• judge_reward, the judge's grade of a proof that was cut off and so scored 0;
• the judge's scores, the completion's length and finish reason, and wall time.

The training proofs themselves are not in this table. Every 10th update's are in the Trackio dashboard's traces.

Provenance

These tables are converted from the run's own logs (rollouts.jsonl, train.log, each eval's Harbor trials), with no filtering. A repeat of the step-100 eval, launched by mistake when the run resumed, is left out; the original step-100 eval is kept. The judge is an LLM grading against QED-Nano's marking schemes. A hand audit found it lenient on proofs that keep contradictions around a correct argument, so read reward as the judge's view, not a verified proof check.