Dataset Card for Omni-MATH-2

Benchmarks are important tools for tracking progress in the development of large language models (LLMs). However, inaccuracies in datasets and evaluation methods often undermine their effectiveness. Here, we present Omni-MATH-2: a manually revised version of the original Omni-MATH dataset that preserves its size (n = 4,428) while significantly improving LaTeX compilability, solvability, and verifiability. A total of 647 problems were edited (14.6%) and 247 were tagged as non-standard (5.6%). This means they contain images, request estimations or proofs, or are degenerate (e.g. duplicate, no reference answer, empty problem etc). We have released multiple evaluation-ready subsets, notably Omni-MATH-2-Filtered (n = 4,181), from which the tagged non-standard questions have been excluded to ensure suitability for judging exact answers. The cleaning process is illustrated in the figure below.

• Paper: https://arxiv.org/abs/2601.19532
• Github Repo: https://github.com/MartheBallon/Benchmarks-saturate-when-the-model-gets-smarter-than-the-judge

Overview of the cleaning process

Uses

from datasets import load_dataset
dataset = load_dataset("martheballon/Omni-MATH-2")

Citation

If you find our dataset helpful, you are welcome to cite our paper.

@misc{ballon2026benchmarkssaturatewhen,
      title={Benchmarks Saturate When The Model Becomes Smarter Than The Judge}, 
      author={Ballon, Marthe and Algaba, Andres and Verbeken, Brecht and Ginis, Vincent},
      year={2026},
      eprint={2601.19532},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2601.19532}, 
}

Acknowledgements

We thank the authors and contributors of the original Omni-MATH dataset. Their release of the benchmark and verifier Omni-Judge enabled the analyses in this paper.

@misc{gao2024omnimathuniversalolympiadlevel,
      title={Omni-MATH: A Universal Olympiad Level Mathematic Benchmark For Large Language Models}, 
      author={Bofei Gao and Feifan Song and Zhe Yang and Zefan Cai and Yibo Miao and Qingxiu Dong and Lei Li and Chenghao Ma and Liang Chen and Runxin Xu and Zhengyang Tang and Benyou Wang and Daoguang Zan and Shanghaoran Quan and Ge Zhang and Lei Sha and Yichang Zhang and Xuancheng Ren and Tianyu Liu and Baobao Chang},
      year={2024},
      eprint={2410.07985},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.07985}, 
}