Open Parallel Think

Parallel reasoning datasets for mathematics and code. Each ZIP contains one JSONL file whose rows group multiple trajectories for a task.

Files

File	Description	Size
open_parallel_think_math.jsonl.zip	Parallel mathematical reasoning trajectories	329.71 MiB
open_parallel_think_code.jsonl.zip	Parallel code reasoning trajectories	595.97 MiB

Dataset statistics

Dataset	Rows	Total trajectories	Trajectory groups
Math	59,247	508,056	63,507
Code	40,000	211,529	35,334

Trajectory groups are counted by unique guid.

Data structure

Each JSONL row has the following structure:

{
  "guid": "string",
  "trajectories": ["string"],
  "queries": [
    {
      "answer": "string",
      "question": "string"
    }
  ],
  "task": "string",
  "statistic": {},
  "num_tokens": 0,
  "frontier_demonstration": "string",
  "source": "string"
}

For Math, statistic contains answer_lengths, correct_ratio, and extracted_answers; Math rows also include accuracy_bucket. For Code, statistic contains answer_lengths, correct_ratio, corrects, and models.

License

The datasets are licensed under the Creative Commons Attribution 4.0 International License. Third-party source material remains subject to any separate rights or notices that apply.

Citation

@misc{open_parallel_think_math_2026,
  author    = {Chen, Zeming and Wu, Haotian and Bosselut, Antoine},
  title     = {Open Parallel Think Math},
  year      = {2026},
  publisher = {EPFL NLP Lab},
  url       = {https://github.com/epfl-nlp/open_parallel_think}
}

@misc{open_parallel_think_code_2026,
  author    = {Chen, Zeming and Wu, Haotian and Bosselut, Antoine},
  title     = {Open Parallel Think Code},
  year      = {2026},
  publisher = {EPFL NLP Lab},
  url       = {https://github.com/epfl-nlp/open_parallel_think}
}