OpenScience

## Dataset Description:

**OpenScience** is a multi-domain synthetic dataset designed to improve general-purpose reasoning in large language models (LLMs). The dataset contains multiple-choice question-answer pairs with detailed reasoning traces and spans across diverse scientific domains, including STEM, law, economics, and humanities. OpenScience aims to boost accuracy on advanced benchmarks such as GPQA-Diamond and MMLU-Pro via supervised finetuning or reinforcement learning.

This dataset is ready for commercial use.

We identify each split using the format *OS-{ModelName}-{ChoiceCount}*.
The \{ModelName\} component in the ID is an abbreviation corresponding to the language model used to generate the questions.
The number indicates whether the split uses four or ten answer choices. The four-choice splits focus on specialized science topics, the ten-choice splits include both science and humanities questions. All splits use [DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1) for solution generation, applying majority voting across multiple samples to establish the reference answer.

All datasets undergo decontamination against the test sets of [MMLU](https://arxiv.org/abs/2009.03300), [MMLU-Pro](https://arxiv.org/abs/2406.01574), and [GPQA](https://arxiv.org/abs/2311.12022), using the [lm-sys pipeline](https://arxiv.org/abs/2311.04850).

Models used for questions generation:
[Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct), [Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct), [Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B)

## Dataset Owner(s):
NVIDIA Corporation

## Dataset Creation Date:
15/05/2025

## License/Terms of Use:
Governing Terms: This dataset is licensed under a Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/legalcode.en).

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:
OpenScience is intended to be used by the community to continue to improve open models. The data may be freely used to train and evaluate.
<br>

## Data Version:
- v1

## Dataset Characterization
Data Collection Method:
*Synthetic <br>

Labeling Method:
*Synthetic <br>

## Dataset Format
Text (.jsonl files with text fields)

## Dataset Quantification
Record Count: 6M question-answer pairs
Total Data Storage: 80GB

## 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/).

**数据集描述：**
**OpenScience** 是一款多领域合成数据集，旨在提升大语言模型（Large Language Model，LLM）的通用推理能力。该数据集包含带有详细推理过程的多项选择题问答对，覆盖STEM、法学、经济学与人文社科等多元科学领域。OpenScience旨在通过监督微调或强化学习，提升模型在GPQA-Diamond、MMLU-Pro等高级基准评测中的准确率。
本数据集支持商业使用。

我们采用*OS-{模型名称}-{选项数量}*的格式对各数据拆分进行标识。其中ID中的{模型名称}字段为生成对应问题所用语言模型的缩写，数字字段则表示该拆分集采用4个还是10个候选答案：4选项拆分聚焦于特定科学主题，10选项拆分同时涵盖科学与人文社科问题。所有拆分集均使用[DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)生成解题过程，并通过多样本多数投票确定参考答案。

所有数据集均通过[lm-sys pipeline](https://arxiv.org/abs/2311.04850)工具，针对[MMLU](https://arxiv.org/abs/2009.03300)、[MMLU-Pro](https://arxiv.org/abs/2406.01574)及[GPQA](https://arxiv.org/abs/2311.12022)的测试集完成了数据去污染处理。

问题生成所用模型包括：
[Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct)、[Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct)、[Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B)

## 数据集所有者
英伟达公司（NVIDIA Corporation）

## 数据集创建日期
2025年5月15日

## 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，https://creativecommons.org/licenses/by/4.0/legalcode.en）进行授权。
本数据集包含使用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进行分发或公开提供，则该人工智能模型需遵守Qwen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
OpenScience旨在供社区用于持续改进开源模型，相关数据可免费用于模型训练与评估。

## 数据版本
- v1

## 数据集特征描述
### 数据收集方式：*合成生成*
### 标注方式：*合成生成*

## 数据集格式
文本格式（.jsonl文件，包含文本字段）

## 数据集量化统计
记录数量：600万条问答对
总数据存储量：80GB

## 伦理考量
英伟达（NVIDIA）认为，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，以支撑各类人工智能应用的开发。
开发者在按照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应用场景的要求，并应对可能出现的产品误用问题。
请通过[此链接](https://www.nvidia.com/en-us/support/submit-security-vulnerability/)提交安全漏洞或英伟达人工智能相关问题。

搜集汇总

背景与挑战

背景概述

OpenScience是一个多领域科学问答合成数据集，包含600万道多选题，涵盖STEM、法律等多个学科，旨在提升大语言模型的推理能力。数据集采用双许可协议，由NVIDIA公司使用Qwen系列模型生成，总大小约80GB。