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

This dataset contains 49K mathematical problems and their answers, sourced from NuminaMath and DeepScaler-Preview, and processed through strict filtering rules that exclude multi-subproblem questions, multiple-choice questions, true/false questions, long and complex answers, proof-based problems, and data involving diagrams. This dataset was used to train the AceReason-Nemotron series of models, which achieved excellent results on mathematical benchmarks such as AIME 2024 and AIME 2025, making it suitable for community deployment of reinforcement learning tasks based on large language models.

Dataset Composition

The dataset primarily consists of mathematical problems and their corresponding answers. The data was filtered to remove types unsuitable for reinforcement learning training, such as questions with multiple subproblems, multiple-choice questions, true/false questions, questions with overly long or complex answers, proof-based problems, and questions involving diagrams. The dataset contains approximately 49K samples, is in English, follows a standard dataset structure, and is suitable for text generation and mathematical reasoning tasks.

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} }