Nemotron-PrismMath

# Nemotron-PrismMath

**Jaehun Jung, Seungju Han\*, Ximing Lu\*, Skyler Hallinan\*, David Acuna, Shrimai Prabhumoye, Mostofa Patwary, Mohammad Shoeybi, Bryan Catanzaro, Yejin Choi**

[Paper](https://arxiv.org/abs/2505.20161) [Project Page](https://nvlabs.github.io/prismatic-synthesis/)

## Dataset Description
<p align="center">
<img style="width:50%" src="prismatic-synthesis.png">
</p>

Nemotron-PrismMath is a state-of-the-art math reasoning dataset with diverse, novel math problems. This dataset is ready for commercial/non-commercial use.

- The dataset consists of 1M math problem-solution pairs generated via **Prismatic Synthesis**, our novel algorithm to generate diverse synthetic
data with novel reasoning patterns. We start from 94k seed samples in [OpenR1-Math](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k), and use
[Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) to generate new problems and
[R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) to generate corresponding solutions.
- **All problems are generated by 32B / 72B LLMs without crawling from the web**,
thus the dataset does not overlap with existing datasets such as [NuminaMath-1.5](https://huggingface.co/datasets/AI-MO/NuminaMath-1.5).
- **Our model trained on the dataset outperforms state-of-the-art SFT models on OOD benchmarks**, making it a useful base model for further RL training of reasoning models.

<p align="center">
<img style="width:80%" src="prismmath-7b-perf.png">
</p>

Check out our [paper]() for more details!

## Citation
If you find our work useful, please consider citing us!

```bibtex
@misc{prismatic-synthesis,
title={Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning},
author={Jaehun Jung and Seungju Han and Ximing Lu and Skyler Hallinan and David Acuna and Shrimai Prabhumoye and Mostafa Patwary and Mohammad Shoeybi and Bryan Catanzaro and Yejin Choi},
year={2025},
eprint={2505.20161},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2505.20161},
}
```

## Dataset Owner(s):
NVIDIA Corporation

## Dataset Creation Date:
April 28, 2025

## 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.

This dataset contains synthetic data created using Qwen2.5-72B-Instruct.
If this dataset is used to create, train, fine tune, or otherwise improve an AI model,
which is distributed or made available, such AI model may be subject to redistribution
and use requirements in the [Qwen License Agreement](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE).

## Intended Usage:
The PrismMath Dataset is intended to be used by the community to deploy reinforcement learning with LLMs.
The data may be freely used to train and evaluate.

## Data Version:
- v1

## 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/).

# Nemotron-PrismMath

**郑在勋, 韩承柱*, 卢熙明*, 斯凯勒·哈利南*, 大卫·阿库纳, 什里迈·普拉布莫耶, 莫斯托法·帕特瓦里, 穆罕默德·舒埃比, 布莱恩·卡塔扎罗, 崔在仁**

[论文](https://arxiv.org/abs/2505.20161) [项目主页](https://nvlabs.github.io/prismatic-synthesis/)

## 数据集说明

<p align="center">
<img style="width:50%" src="prismatic-synthesis.png">
</p>

Nemotron-PrismMath是一款顶尖的数学推理数据集，涵盖多样化且新颖的数学问题。本数据集可用于商业及非商业用途。

- 本数据集包含100万道数学问题-解答配对数据，通过**Prismatic Synthesis（棱镜合成算法）**生成，该算法为我们提出的全新方案，可生成具备新颖推理模式的多样化合成数据。我们以[OpenR1-Math](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k)中的9.4万条种子样本为起点，使用[Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct)生成新的数学问题，并使用[R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)生成对应的解答。
- **所有问题均由320亿/720亿参数的大语言模型（LLM）生成，未通过网络爬取获取**，因此本数据集与[NuminaMath-1.5](https://huggingface.co/datasets/AI-MO/NuminaMath-1.5)等现有数据集无重叠。
- **基于本数据集训练的模型在分布外（OOD）基准测试中优于现有顶尖监督微调（SFT）模型**，可作为推理模型进一步强化学习训练的优质基础模型。

<p align="center">
<img style="width:80%" src="prismmath-7b-perf.png">
</p>

如需了解更多细节，请查阅我们的[论文]()！

## 引用格式

若您认为本工作对您有帮助，请考虑引用我们的成果：

bibtex
@misc{prismatic-synthesis,
title={Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning},
author={Jaehun Jung and Seungju Han and Ximing Lu and Skyler Hallinan and David Acuna and Shrimai Prabhumoye and Mostafa Patwary and Mohammad Shoeybi and Bryan Catanzaro and Yejin Choi},
year={2025},
eprint={2505.20161},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2505.20161},
}

## 数据集所有者：

英伟达（NVIDIA Corporation）

## 数据集创建日期：

2025年4月28日

## 许可与使用条款：

### 许可条款
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，简称CC BY 4.0）进行授权，许可协议详情可访问 https://creativecommons.org/licenses/by/4.0/legalcode。

本数据集包含使用[Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct)生成的合成数据。若使用本数据集创建、训练、微调或以其他方式改进人工智能模型，并对该模型进行分发或公开提供，则该人工智能模型需遵守[通义千问许可协议](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE)中的重新分发与使用要求。

## 预期用途：

本PrismMath数据集旨在供社区用于开展大语言模型的强化学习部署，可自由用于模型训练与评估工作。

## 数据版本：

- v1

## 伦理考量：

英伟达坚信，可信人工智能是一项共同责任，我们已制定相关政策与实践规范，以支持各类人工智能应用的开发。开发者在按照本服务条款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使用场景的要求，并防范可能出现的产品误用问题。

若发现安全漏洞或有英伟达人工智能相关问题，请[在此提交](https://www.nvidia.com/en-us/support/submit-security-vulnerability/)。