ReaRAG-20k

# 📘 Dataset Card for ReaRAG-20k

<p align="center">
🤗 <a href="https://huggingface.co/THU-KEG/ReaRAG-9B" target="_blank">Model</a> • 💻 <a href="https://github.com/THU-KEG/ReaRAG" target="_blank">GitHub</a> • 📃 <a href="https://arxiv.org/abs/2503.21729" target="_blank">Paper</a>
</p>

ReaRAG-20k is a reasoning-focused dataset designed for training the ReaRAG model. It contains approximately 20,000 multi-turn retrieval examples constructed from the QA datasets such as HotpotQA, MuSiQue, and Natural Questions (NQ).

Each instance follows a conversational format supporting reasoning and retrieval steps:
```json
{
"messages": [{"role": "user", "content": "..."},
{"role": "assistant", "reasoning": "..."},
{"role": "observation", "content": "..."}, ...]
}
```
During sft, the loss is computed only on messages that contain the `reasoning` key, rather than the `content` key.

# 🔗 Resources

- **Code Repository:** [💻 GitHub](https://github.com/THU-KEG/ReaRAG)
- **Paper:** [📃 ArXiv](https://arxiv.org/abs/2503.21729)
- **Model:** [🤗 Huggingface](https://huggingface.co/THU-KEG/ReaRAG-9B). A model based on GLM-4-9B, sft on this dataset.

# 📚 Citation
If you use this dataset in your research or projects, please consider citing our work:
```
@article{lee2025rearag,
title={ReaRAG: Knowledge-guided Reasoning Enhances Factuality of Large Reasoning Models with Iterative Retrieval Augmented Generation},
author={Lee, Zhicheng and Cao, Shulin and Liu, Jinxin and Zhang, Jiajie and Liu, Weichuan and Che, Xiaoyin and Hou, Lei and Li, Juanzi},
journal={arXiv preprint arXiv:2503.21729},
year={2025}
}
```

# 📘 ReaRAG-20k 数据集卡片

<p align="center">
🤗 <a href="https://huggingface.co/THU-KEG/ReaRAG-9B" target="_blank">模型</a> • 💻 <a href="https://github.com/THU-KEG/ReaRAG" target="_blank">GitHub仓库</a> • 📃 <a href="https://arxiv.org/abs/2503.21729" target="_blank">研究论文</a>
</p>

ReaRAG-20k是一款聚焦推理任务的专用数据集，旨在为ReaRAG模型的训练提供支撑。该数据集包含约20000条多轮检索示例，其构建数据源涵盖HotpotQA、MuSiQue以及Natural Questions（NQ）等主流问答数据集。

每个数据实例均采用适配推理与检索流程的对话格式，具体结构如下：
json
{
"messages": [{"role": "user", "content": "..."},
{"role": "assistant", "reasoning": "..."},
{"role": "observation", "content": "..."}, ...]
}

在监督微调（Supervised Fine-Tuning，SFT）阶段，仅对带有`reasoning`字段的对话消息计算损失，而非带有`content`字段的消息。

# 🔗 相关资源

- **代码仓库：** [💻 GitHub](https://github.com/THU-KEG/ReaRAG)
- **研究论文：** [📃 ArXiv](https://arxiv.org/abs/2503.21729)
- **预训练模型：** [🤗 Hugging Face](https://huggingface.co/THU-KEG/ReaRAG-9B)。该模型基于GLM-4-9B构建，并使用本数据集完成监督微调。

# 📚 引用规范

若您在研究工作或实际项目中使用本数据集，请引用如下文献：

@article{lee2025rearag,
title={ReaRAG: Knowledge-guided Reasoning Enhances Factuality of Large Reasoning Models with Iterative Retrieval Augmented Generation},
author={Lee, Zhicheng and Cao, Shulin and Liu, Jinxin and Zhang, Jiajie and Liu, Weichuan and Che, Xiaoyin and Hou, Lei and Li, Juanzi},
journal={arXiv preprint arXiv:2503.21729},
year={2025}
}