Datasets:

emendes3
/

e1-verifiable

e1-verifiable contains challenging problems from the American Invitational Mathematics Examination (AIME), i.e., pass@32 = 0 for Qwen2.5-7B-Instruct. The solutions were sourced from Yue et al.'s scrape of the AoPS Wiki.

This dataset was collected in this work: Making Expert Reasoning Learnable with Self-Distillation. If you use this dataset, please cite this paper:

@inproceedings{
    mendes2026selfdistill,
    title={Making Expert Reasoning Learnable with Self-Distillation},
    author={Mendes, Ethan and Park, Jungsoo and Ritter, Alan},
    booktitle={Forty-third International Conference on Machine Learning},
    year={2026},
    url={https://openreview.net/forum?id=JG6f02X29a}
}