Ring-lite-sft-data

<p align="center">
<img src="https://modelscope.cn/models/inclusionAI/Ring-lite/resolve/master/ant-bailing.png" width="100"/>
<p>

<p align="center">
🤖 <a href="https://modelscope.cn/organization/inclusionAI">ModelScope</a>
🤗 <a href="https://huggingface.co/inclusionAI">HuggingFace</a>
🖥️ <a href="https://github.com/inclusionAI/Ring">GitHub</a>
<p>

## Ring-lite-sft-data

This is a the SFT data used during the fine-tuning of the [Ring-lite](https://modelscope.cn/models/inclusionAI/Ring-lite/) model. The query pool was sourced from open-source repositories and further enriched through synthetic generation using large language models (LLMs). To ensure the production of high-fidelity responses with Long-CoT, we implemented an iterative refinement pipeline that synergistically combines automated model generation, expert manual annotation, and rejection sampling mechanisms. After that, rigorous data-cleansing protocols were applied, including detection and removal of repetitive patterns, mixed-language artifacts, and other noise sources, to yield a robust and high-quality dataset.
The final data is predominantly dominated by three major domains: Mathematics (64.5\%), Code (25.5\%), and Science (9.2\%). The remaining portion of the dataset includes contributions from other categories, such as medicine and history domains.

More details are reported in our [technical report](https://arxiv.org/abs/2506.14731).

**Note**: Only a partial subset of the complete dataset is publicly released due to third-party data licensing restrictions and procurement agreements. The published portion has been carefully selected to comply with all copyright requirements while maintaining research utility.

## Citation
```
@misc{ringteam2025ringlitescalablereasoningc3postabilized,
title={Ring-lite: Scalable Reasoning via C3PO-Stabilized Reinforcement Learning for LLMs},
author={Ling Team},
year={2025},
eprint={2506.14731},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.14731},
}
```

<p align="center">
<img src="https://modelscope.cn/models/inclusionAI/Ring-lite/resolve/master/ant-bailing.png" width="100"/>
</p>

<p align="center">
🤖 <a href="https://modelscope.cn/organization/inclusionAI">魔搭社区（ModelScope）</a>
🤗 <a href="https://huggingface.co/inclusionAI">HuggingFace</a>
🖥️ <a href="https://github.com/inclusionAI/Ring">GitHub</a>
</p>

## Ring-lite微调数据集（Ring-lite-sft-data）

本数据集为[Ring-lite](https://modelscope.cn/models/inclusionAI/Ring-lite/)模型微调阶段所用的监督微调（Supervised Fine-Tuning，SFT）数据。查询池源自各类开源仓库，并借助大语言模型（LLMs）完成合成数据生成以进一步扩充样本规模。为生成符合长链式思维（Long-Chain of Thought，Long-CoT）规范的高保真回复，我们搭建了迭代优化流水线，协同整合了自动化模型生成、专家人工标注与拒绝采样三大机制。随后我们执行了严格的数据清洗流程，包括检测并移除重复模式、混合语言伪迹及其他各类噪声，最终得到鲁棒性强、质量上乘的数据集。

最终数据集主要涵盖三大核心领域：数学（64.5%）、代码（25.5%）与科学（9.2%），数据集剩余部分则涵盖医学、历史等其他细分领域。

更多细节可参阅我们发布的[技术报告](https://arxiv.org/abs/2506.14731)。

**注意**：由于第三方数据许可限制与采购协议约束，本次仅公开完整数据集的部分子集。公开子集经过精心筛选，既符合所有版权要求，又保留了科研应用价值。

## 引用

@misc{ringteam2025ringlitescalablereasoningc3postabilized,
title={Ring-lite: 基于C3PO稳定强化学习的大语言模型可扩展推理方法},
author={Ling Team},
year={2025},
eprint={2506.14731},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.14731},
}