Sangsang/rewind-rl-mixed453

Rewind RL: 453 mixed-success math problems

The exact 453 question/answer pairs used in the current rewind RL experiment. train.jsonl preserves the training file's contents and row order. It contains only question and answer (both strings); answers are reference final answers. Training generates fresh trajectories on policy. No generated reasoning, teacher diagnoses, hints, or checkpoints are included in this dataset.

Selection

Start with 1,024 candidate problems from the English training subset of open-r1/DAPO-Math-17k-Processed. The local pool removed duplicate questions and screened AIME24/25 overlap using normalized word 8-grams (intersection at least half the smaller set).

Generate eight attempts per problem using hbin0701/Qwen3-1.7B-rewind-sftv3-wait, with thinking enabled, temperature 1.0, top-p 1.0, and a 16,384-token response limit. The collection prompt contains rewind instructions, but the rewind token is blocked throughout these attempts. Keep every problem with 1–7 correct answers out of 8. This count measures empirical success frequency, rather than the binary event of at least one success often called pass@8.

Correct out of 8	Problems
1	50
2	43
3	50
4	44
5	64
6	67
7	135
Total	453

All 453 problems belong to the train split. Selection measures difficulty for this particular frozen model; it does not establish that a rewind will help at any particular reasoning position. selection.json records each row's success count, provenance, and the exact training-file SHA-256.

Attribution

Derived from DAPO-Math-17k-Processed, itself processed from BytedTsinghua-SIA/DAPO-Math-17k. Please credit the upstream datasets and the DAPO paper. This subset does not grant additional rights to the underlying problem text; refer to the upstream sources for terms.