Skywork-OR1-RL-Data

<div align="center">

# 🤔 Skywork-OR1-RL-Data

</div>

<div align="center">

[](https://huggingface.co/collections/Skywork/skywork-or1-67fa1bcb41b436ef2def76b9)
[](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data)
[](https://github.com/SkyworkAI/Skywork-OR1)
[](https://capricious-hydrogen-41c.notion.site/Skywork-Open-Reaonser-Series-1d0bc9ae823a80459b46c149e4f51680)

[](https://github.com/SkyworkAI/Skywork-OR1/stargazers)
[](https://github.com/SkyworkAI/Skywork-OR1/fork)

</div>

## 🔥 News

- **April 15, 2025**: We are excited to release our RL training dataset [`Skywork-OR1-RL-Data`](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data)
- For our final training phase, we filtered problems based on their difficulty levels (0-16, higher values indicate harder problems) relative to specific model variants (DeepSeek-R1-Distill-Qwen-{1.5,7,32}B. For each model variant, we excluded problems with difficulty values of 0 and 16 specific to that model from its training data.
- You can check our [Skywork-OR1](https://github.com/SkyworkAI/Skywork-OR1?tab=readme-ov-file#training-data-preparation) repository for training data preparation steps.
- **Note**: Due to an accidental early release, a version with incorrect difficulty fields was briefly public. Please make sure to use either the newest version (recommended) or any version at this [commit](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data/commit/b48ac2ee70ae3dc5d6db769f232e8a966cb89240) and after.

## 📖 Overview

[`Skywork-OR1-RL-Data`](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data) is **a dataset of verifiable, challenging, and diverse math problems (105K) and coding questions (14K)**. This dataset is used to train the **`Skywork-OR1`** (Open Reasoner 1) model series, which consists of powerful math and code reasoning models trained using large-scale rule-based reinforcement learning with carefully designed datasets and training recipes. This series includes two general-purpose reasoning modelsl, **`Skywork-OR1-7B-Preview`** and **`Skywork-OR1-32B-Preview`**, along with a math-specialized model, **`Skywork-OR1-Math-7B`**.

- **[`Skywork-OR1-Math-7B`](https://huggingface.co/Skywork/Skywork-OR1-Math-7B)** is specifically optimized for mathematical reasoning, scoring **69.8** on AIME24 and **52.3** on AIME25 — well ahead of all models of similar size.
- **[`Skywork-OR1-32B-Preview`](https://huggingface.co/Skywork/Skywork-OR1-32B-Preview)** delivers the 671B-parameter Deepseek-R1 performance on math tasks (AIME24 and AIME25) and coding tasks (LiveCodeBench).
- **[`Skywork-OR1-7B-Preview`](https://huggingface.co/Skywork/Skywork-OR1-7B-Preview)** outperforms all similarly sized models in both math and coding scenarios.

We select, clean, and curate math and coding problems from open-source datasets, including

- [NuminaMath-1.5](https://huggingface.co/datasets/AI-MO/NuminaMath-1.5)
- [DeepScaleR-Preview-Dataset](https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset)
- [STILL-3-Preview-RL-Data](https://huggingface.co/datasets/RUC-AIBOX/STILL-3-Preview-RL-Data)
- [Omni-Math](https://huggingface.co/datasets/KbsdJames/Omni-MATH)
- [AIME problems prior to 2024](https://huggingface.co/datasets/gneubig/aime-1983-2024)
- [LeetCodeDataset](https://huggingface.co/datasets/newfacade/LeetCodeDataset)
- [TACO](https://huggingface.co/datasets/BAAI/TACO)

We conduct **model-aware difficulty estimation** for each problem and model and conduct **rigorous quality assessment prior to training** via both human and LLM-as-a-Judge to ensure training efficiency and effectiveness. We also perform deduplication within the dataset and remove similar problems from AIME 24, AIME 25, and LiveCodeBench to prevent data contamination.

## 📄 Technical Report

Our technical report will be released soon. Stay tuned!

## 📚 Citation

Please cite the following:

```bibtex
@article{he2025skywork,
title={Skywork Open Reasoner 1 Technical Report},
author={He, Jujie and Liu, Jiacai and Liu, Chris Yuhao and Yan, Rui and Wang, Chaojie and Cheng, Peng and Zhang, Xiaoyu and Zhang, Fuxiang and Xu, Jiacheng and Shen, Wei and Li, Siyuan and Zeng, Liang and Wei, Tianwen and Cheng, Cheng and An, Bo and Liu, Yang and Zhou, Yahui},
journal={arXiv preprint arXiv:2505.22312},
year={2025}
}

@misc{skywork-or1-2025,
title={Skywork Open Reasoner Series},
author = {He, Jujie and Liu, Jiacai and Liu, Chris Yuhao and Yan, Rui and Wang, Chaojie and Cheng, Peng and Zhang, Xiaoyu and Zhang, Fuxiang and Xu, Jiacheng and Shen, Wei and Li, Siyuan and Zeng, Liang and Wei, Tianwen and Cheng, Cheng and Liu, Yang and Zhou, Yahui},
howpublished={\url{https://capricious-hydrogen-41c.notion.site/Skywork-Open-Reaonser-Series-1d0bc9ae823a80459b46c149e4f51680}},
note={Notion Blog},
year={2025}
}
```

<div align="center">

# 🤔 Skywork-OR1-RL-Data 数据集

</div>

<div align="center">

[](https://huggingface.co/collections/Skywork/skywork-or1-67fa1bcb41b436ef2def76b9)
[](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data)
[](https://github.com/SkyworkAI/Skywork-OR1)
[](https://capricious-hydrogen-41c.notion.site/Skywork-Open-Reaonser-Series-1d0bc9ae823a80459b46c149e4f51680)

[](https://github.com/SkyworkAI/Skywork-OR1/stargazers)
[](https://github.com/SkyworkAI/Skywork-OR1/fork)

</div>

## 🔥 最新动态

- **2025年4月15日**：我们正式发布强化学习（Reinforcement Learning，RL）训练数据集 [`Skywork-OR1-RL-Data`](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data)
- 在最终训练阶段，我们会针对特定模型变体（DeepSeek-R1-Distill-Qwen-{1.5,7,32}B）的难度等级（取值范围0-16，数值越高代表题目难度越大）筛选题目。针对每个模型变体，我们会从其训练数据中排除该模型专属的难度值为0和16的题目。
- 你可以前往我们的 [`Skywork-OR1`](https://github.com/SkyworkAI/Skywork-OR1?tab=readme-ov-file#training-data-preparation) 仓库查看训练数据的准备步骤。
- **注意**：由于意外提前发布，存在错误难度字段的版本曾短暂公开。请务必使用最新版本（推荐）或该 [提交记录](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data/commit/b48ac2ee70ae3dc5d6db769f232e8a966cb89240) 及之后的版本。

## 📖 数据集概览

[`Skywork-OR1-RL-Data`](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data) 是**包含10.5万个可验证、具挑战性且多样化的数学问题与1.4万个编码问题的数据集**。本数据集用于训练**`Skywork-OR1`（开放推理器1，Open Reasoner 1）**模型系列，该系列是通过大规模基于规则的强化学习结合精心设计的数据集与训练配方训练得到的高性能数学与代码推理模型。该系列包含两款通用推理模型：**`Skywork-OR1-7B-Preview`**与**`Skywork-OR1-32B-Preview`**，以及一款数学专用模型**`Skywork-OR1-Math-7B`**。

- **[`Skywork-OR1-Math-7B`](https://huggingface.co/Skywork/Skywork-OR1-Math-7B)** 专门针对数学推理优化，在**美国数学邀请赛（American Invitational Mathematics Examination，AIME）2024**上得分69.8，在AIME2025上得分52.3，远超同尺寸所有其他模型。
- **[`Skywork-OR1-32B-Preview`](https://huggingface.co/Skywork/Skywork-OR1-32B-Preview)** 在数学任务（AIME24与AIME25）和编码任务（LiveCodeBench）上达到671B参数Deepseek-R1的性能表现。
- **[`Skywork-OR1-7B-Preview`](https://huggingface.co/Skywork/Skywork-OR1-7B-Preview)** 在数学与编码场景下均优于同尺寸所有其他模型。

我们从多个开源数据集选取、清洗并整理了数学与编码问题，包括：
- [NuminaMath-1.5](https://huggingface.co/datasets/AI-MO/NuminaMath-1.5)
- [DeepScaleR-Preview-Dataset](https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset)
- [STILL-3-Preview-RL-Data](https://huggingface.co/datasets/RUC-AIBOX/STILL-3-Preview-RL-Data)
- [Omni-Math](https://huggingface.co/datasets/KbsdJames/Omni-MATH)
- [2024年前的AIME题目](https://huggingface.co/datasets/gneubig/aime-1983-2024)
- [力扣（LeetCode）数据集](https://huggingface.co/datasets/newfacade/LeetCodeDataset)
- [TACO](https://huggingface.co/datasets/BAAI/TACO)

我们针对每个问题与模型开展**模型感知的难度估计**，并在训练前通过人工与大语言模型作为评判者进行严格的质量评估，以确保训练效率与效果。同时我们在数据集内部进行去重，并移除AIME24、AIME25与LiveCodeBench中的相似题目，防止数据污染。

## 📄 技术报告

我们的技术报告即将发布，敬请期待！

## 📚 引用格式

请使用以下引用格式：

bibtex
@article{he2025skywork,
title={Skywork Open Reasoner 1 Technical Report},
author={He, Jujie and Liu, Jiacai and Liu, Chris Yuhao and Yan, Rui and Wang, Chaojie and Cheng, Peng and Zhang, Xiaoyu and Zhang, Fuxiang and Xu, Jiacheng and Shen, Wei and Li, Siyuan and Zeng, Liang and Wei, Tianwen and Cheng, Cheng and An, Bo and Liu, Yang and Zhou, Yahui},
journal={arXiv preprint arXiv:2505.22312},
year={2025}
}

@misc{skywork-or1-2025,
title={Skywork Open Reasoner Series},
author = {He, Jujie and Liu, Jiacai and Liu, Chris Yuhao and Yan, Rui and Wang, Chaojie and Cheng, Peng and Zhang, Xiaoyu and Zhang, Fuxiang and Xu, Jiacheng and Shen, Wei and Li, Siyuan and Zeng, Liang and Wei, Tianwen and Cheng, Cheng and Liu, Yang and Zhou, Yahui},
howpublished={url{https://capricious-hydrogen-41c.notion.site/Skywork-Open-Reaonser-Series-1d0bc9ae823a80459b46c149e4f51680}},
note={Notion 博客},
year={2025}
}