AceReason-Math is a high-quality, verifiable, challenging, and diverse mathematical dataset released by NVIDIA in 2025, with related paper results at AceReason-Nemotron: Advancing Math and Code Reasoning through Reinforcement Learning, aimed at training math reasoning models through reinforcement learning.

The dataset comprises 49K math problems along with their answers, sourced from NuminaMath and DeepScaler-Preview, and processed using strict filtering rules that exclude multi-part questions, multiple-choice items, true/false questions, long/complex answers, proof-based problems, and data involving charts/diagrams. This dataset was used to train the AceReason-Nemotron series of models, which achieved outstanding performance on benchmarks like AIME 2024 and AIME 2025, making it suitable for community deployment of LLM-based reinforcement learning tasks.

Dataset Composition

This dataset primarily consists of mathematics problems paired with corresponding solutions. Data underwent screening to filter out types unsuitable for reinforcement learning training, including those containing sub-problems, multiple-choice formats, yes/no queries, excessively lengthy or complex responses, proofs, and visual/chart elements. With approximately 49K samples available exclusively in English format structured according standard conventions, this resource supports both textual generation capabilities alongside advanced problem-solving applications requiring logical deduction skills necessary within academic contexts today!

<div style="text-align:center">
<img src="https://hyperai.cn-bj.ufileos.com/d6166f58-1a35-4c07-94dc-fa399dd7e21c" alt="数据集示例">
</div>

## Citation

```bibtex
@article{chen2025acereason,
  title={AceReason-Nemotron: Advancing Math and Code Reasoning through Reinforcement Learning},
  author={Chen, Yang and Yang, Zhuolin and Liu, Zihan and Lee, Chankyu and Xu, Peng and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei},
  journal={arXiv preprint arXiv:2505.16400},
  year={2025}
}