Qwen3-32B Math n32 16384 (44k Queries)

This dataset contains multi-sampled rollout traces from Qwen3-32B on around 44k math queries. For each query, the model is rolled out 32 times with a maximum generation length of 16384 tokens.

Each response is annotated with answer correctness (acc_reward), and includes token-level statistics (action_entropy, action_log_probs) for further analysis and research.

Resources

• Paper: Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability
• Code: Official GitHub Repository
• Hugging Face Collection: rethink-sft-generalization

Highlights

• Source model: Qwen3-32B
• Query scale: ~44k math queries
• Sampling per query: 32 responses
• Max generation length: 16384
• Per-response supervision includes acc_reward (correctness), action_entropy (token-level entropy), and action_log_probs (token-level log-probability).

Generation Details

• Generation model: Qwen3-32B with thinking enabled.
• Maximum response length: 16384.
• Decoding setup: temperature=0.6, top_p=0.95, top_k=20, min_p=0.

File Layout

The dataset is sharded into 64 parquet files:

• qwen3-32b_n32_16384_dp0.parquet
• qwen3-32b_n32_16384_dp1.parquet
• ...
• qwen3-32b_n32_16384_dp63.parquet

Practical Note [Important]

This dataset is very large. Please avoid loading all shards into memory at once. For most workflows, prefer shard-wise iteration, streaming, or memory-mapped parquet/arrow processing.

File Structure

- data_source: string
- question: string
- answer: string
- message: list<element: struct<content: string, role: string>>
- prompt: string
- responses: list<element: struct<
    acc_reward: double,
    action_entropy: list<element: double>,
    action_log_probs: list<element: double>,
    idx: int64,
    length: int64,
    response: string
  >>

Field Notes

• question: input math query.
• answer: reference answer associated with the query.
• message / prompt: chat-style and prompt-style input representations.
• responses: list of sampled responses for the same query.
• responses[*].acc_reward: 1.0 indicates answer judged correct, 0.0 indicates answer judged incorrect.
• responses[*].action_entropy: per-token entropy values.
• responses[*].action_log_probs: per-token log-probability values.
• responses[*].length: generated response length.

Intended Use

For example:

• SFT or offline RL research
• Reasoning data quality analysis
• Calibration and uncertainty studies using entropy/logprob signals
• Diversity of multi-sampled math reasoning traces

Citation

@article{ren2026rethinking_sft_generalization,
  title={Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability},
  author={Qihan Ren and Peng Wang and Ruikun Cai and Shuai Shao and Dadi Guo and Yuejin Xie and Yafu Li and Quanshi Zhang and Xia Hu and Jing Shao and Dongrui Liu},
  journal={arXiv preprint arXiv:2604.06628},
  year={2026}
}