AceReason-Math

# AceReason-Math Dataset

<p align="center">

[](https://arxiv.org/abs/2505.16400)
[](https://huggingface.co/datasets/nvidia/AceReason-Math)
[](https://huggingface.co/collections/nvidia/acereason-682f4e1261dc22f697fd1485)
[](https://huggingface.co/nvidia/AceReason-Nemotron-14B/blob/main/README_EVALUATION.md)
[](https://huggingface.co/papers/2506.13284)
</p>

<img src="fig/main_fig.png" alt="main_fig" style="width: 600px; max-width: 100%;" />

# Overview

AceReason-Math is a high quality, verfiable, challenging and diverse math dataset for training math reasoning model using reinforcement leraning. This dataset contains
- 49K math problems and answer sourced from [NuminaMath](https://huggingface.co/datasets/AI-MO/NuminaMath-1.5) and [DeepScaler-Preview](https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset)
- applying filtering rules to exclude unsuitable data (e.g., multiple sub-questions, multiple-choice, true/false, long and complex answers, proof, figure)
- this dataset was used to train [AceReason-Nemotron](https://huggingface.co/collections/nvidia/acereason-682f4e1261dc22f697fd1485) models, which achieve strong results on math benchmark such as AIME24 and AIME25.

| **Model** | **AIME 2024<br>(avg@64)** | **AIME 2025<br>(avg@64)** |
| :---: | :---: | :---: |
| <small>QwQ-32B</small> | 79.5 | 65.8 |
| <small>DeepSeek-R1-671B</small> | 79.8 | 70.0 |
| <small>Llama-Nemotron-Ultra-253B</small> | 80.8 | 72.5 |
| <small>o3-mini (medium)</small> | 79.6 | 76.7 |
| <small>Light-R1-14B</small> | 74 | 60.2 |
| <small>OpenMath-Nemotron-14B</small> | 76.3 | 63.0 |
| <small>Llama-Nemotron-Super-49B-v1</small> | 67.5 | 60.0 |
| <small>DeepSeek-R1-Distilled-Qwen-14B</small> | 69.7 | 50.2 |
| <small>DeepSeek-R1-Distilled-Qwen-32B</small> | 72.6 | 54.9 |
| [AceReason-Nemotron-7B 🤗](https://huggingface.co/nvidia/AceReason-Nemotron-7B)| 69.0 | 53.6 |
| [AceReason-Nemotron-14B 🤗](https://huggingface.co/nvidia/AceReason-Nemotron-14B)| 78.6 | 67.4 |

## Correspondence to
Yang Chen ([e-mail]), Zhuolin Yang ([e-mail]), Zihan Liu ([e-mail]), Chankyu Lee ([e-mail]), Wei Ping ([e-mail])

### License/Terms of Use:
Governing Terms: This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0)
available at https://creativecommons.org/licenses/by/4.0/legalcode.

### Data Developer:
NVIDIA

### Intended Usage: <br>
The AceReason-Math Dataset is intended to be used by the community to deploy reinforcement learning with LLMs. The data may be used to train and evaluate.<br>

### Release Date: <br>
6/2/2025

### Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).

## Citation
```
@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}
}
```

# AceReason-Math 数据集

<p align="center">

[](https://arxiv.org/abs/2505.16400)
[](https://huggingface.co/datasets/nvidia/AceReason-Math)
[](https://huggingface.co/collections/nvidia/acereason-682f4e1261dc22f697fd1485)
[](https://huggingface.co/nvidia/AceReason-Nemotron-14B/blob/main/README_EVALUATION.md)
[](https://huggingface.co/papers/2506.13284)
</p>

<img src="fig/main_fig.png" alt="主图" style="width: 600px; max-width: 100%;" />

# 概述

AceReason-Math 是一款高质量、可验证、兼具挑战性与多样性的数学数据集，用于通过强化学习（Reinforcement Learning, RL）训练数学推理模型。该数据集包含以下内容：
- 从[NuminaMath](https://huggingface.co/datasets/AI-MO/NuminaMath-1.5)与[DeepScaler-Preview](https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset)中采集的4.9万道数学题目与标准答案
- 通过筛选规则剔除了不适用的数据（例如包含多子问题、选择题、判断题、冗长复杂答案、证明题及配图的题目）
- 该数据集被用于训练[AceReason-Nemotron](https://huggingface.co/collections/nvidia/acereason-682f4e1261dc22f697fd1485)系列模型，该系列模型在AIME2024与AIME2025等数学基准测试中取得了优异成绩。

| **模型** | **AIME 2024<br>(64次采样下平均分)** | **AIME 2025<br>(64次采样下平均分)** |
| :---: | :---: | :---: |
| <small>QwQ-32B</small> | 79.5 | 65.8 |
| <small>DeepSeek-R1-671B</small> | 79.8 | 70.0 |
| <small>Llama-Nemotron-Ultra-253B</small> | 80.8 | 72.5 |
| <small>o3-mini (medium)</small> | 79.6 | 76.7 |
| <small>Light-R1-14B</small> | 74 | 60.2 |
| <small>OpenMath-Nemotron-14B</small> | 76.3 | 63.0 |
| <small>Llama-Nemotron-Super-49B-v1</small> | 67.5 | 60.0 |
| <small>DeepSeek-R1-Distilled-Qwen-14B</small> | 69.7 | 50.2 |
| <small>DeepSeek-R1-Distilled-Qwen-32B</small> | 72.6 | 54.9 |
| [AceReason-Nemotron-7B 🤗](https://huggingface.co/nvidia/AceReason-Nemotron-7B)| 69.0 | 53.6 |
| [AceReason-Nemotron-14B 🤗](https://huggingface.co/nvidia/AceReason-Nemotron-14B)| 78.6 | 67.4 |

## 通讯联系人
Yang Chen ([e-mail]), Zhuolin Yang ([e-mail]), Zihan Liu ([e-mail]), Chankyu Lee ([e-mail]), Wei Ping ([e-mail])

### 许可与使用条款：
本数据集采用知识共享署名4.0国际许可协议（Creative Commons Attribution 4.0 International License, CC BY 4.0）进行授权，协议详情请见https://creativecommons.org/licenses/by/4.0/legalcode。

### 数据开发方：
NVIDIA

### 预期用途：
AceReason-Math 数据集旨在供社区开展基于大语言模型（Large Language Model, LLM）的强化学习研究，可用于模型的训练与评估。

### 发布日期：
2025年6月2日

### 伦理考量：
NVIDIA坚信可信人工智能是一项共同责任，我们已建立相关政策与实践规范，以支持各类人工智能应用的开发。开发者在遵循本服务条款的前提下下载或使用本数据集时，应与其内部模型团队协作，确保该模型符合相关行业与应用场景的要求，并防范可能出现的产品误用。

请通过[此链接](https://www.nvidia.com/en-us/support/submit-security-vulnerability/)报告安全漏洞或NVIDIA人工智能相关问题。

## 引用格式

@article{chen2025acereason,
title={AceReason-Nemotron: 通过强化学习推进数学与代码推理能力},
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预印本 arXiv:2505.16400},
year={2025}
}