Light-IF-SFTData
收藏魔搭社区2025-12-05 更新2025-12-06 收录
下载链接：
https://modelscope.cn/datasets/360zhinao/Light-IF-SFTData
下载链接
链接失效反馈官方服务：
资源简介：
<!-- markdownlint-disable first-line-h1 -->
<!-- markdownlint-disable html -->
<!-- markdownlint-disable no-duplicate-header -->
# Light-IF: Endowing LLMs with Generalizable Reasoning via Preview and Self-Checking
<div align="center">
<img src="https://cdn-uploads.huggingface.co/production/uploads/64eeb81ad0ceda46832e0160/b2_eQV04B8xSdYJZnB2FD.png" width="95%" alt="Light-IF-32B" />
</div>
<hr>
<div align="center">
[](https://arxiv.org/abs/2508.03178)
[](https://huggingface.co/collections/qihoo360/light-if-688753563c174e9d14d6978f)
[](https://zhuanlan.zhihu.com/p/1936535948360918628)
[](https://opensource.org/licenses/Apache-2.0)
</div>
> Here are the cold start data we used to train [Light-IF-32B](https://huggingface.co/qihoo360/Light-IF-32B).
Simply refer to `cold-start.json`
## 🧪 Benchmarks
|Model|SuperClue|IFEval|CFBench|IFBench|
| ---- | ---- | ---- | ---- | ---- |
|Qwen3-4B|0.225|0.888|0.787|0.382|
|Qwen3-8B|0.225|0.888|0.813|0.417|
|Qwen3-32B|0.234|0.877|0.823|0.384|
|Qwen3-235B-A22B|0.244|0.882|0.834|0.423|
|Qwen3-235B-A22B-Thinking-2507|0.434|0.916|0.843|0.475|
|DeepSeek-R1-0528|0.436|0.863|0.827|0.415|
|Doubao-seed-1-6-thinking-250615|0.362|0.832|0.82|0.477|
|Doubao-seed-1-6-thinking-250715|0.345|0.856|0.84|0.366|
|ChatGPT-4o-latest|0.260|0.836|0.807|0.365|
|Deepseek-v3-250324|0.306|0.859|0.833|0.405|
|Doubao-1.5-pro-32k-250115|0.285|0.889|0.797|0.375|
|Kimi-K2|0.227|0.921|0.820|0.395|
|GLM-4.5|0.395|0.893|0.833|0.466|
| [**Light-IF-4B (ours)** 🤗](https://huggingface.co/qihoo360/Light-IF-4B) |**0.445**|**0.916**|**0.80**|**0.443**|
| [**Light-IF-8B (ours)** 🤗](https://huggingface.co/qihoo360/Light-IF-8B) |**0.471**|**0.932**|**0.82**|**0.473**|
| [**Light-IF-32B (ours)** 🤗](https://huggingface.co/qihoo360/Light-IF-32B) |**0.575**|**0.938**|**0.85**|**0.575**|
## Citation
```bibtex
@misc{lightifproj,
title={Light-IF: Endowing LLMs with Generalizable Reasoning via Preview and Self-Checking for Complex Instruction Following},
author={Chenyang Wang, Liang Wen, Shousheng Jia, Xiangzheng Zhang, Liang Xu},
year={2025},
eprint={2508.03178},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2508.03178},
}
```
# Light-IF：通过预览与自检赋予大语言模型（LLMs）通用推理能力
<div align="center">
<img src="https://cdn-uploads.huggingface.co/production/uploads/64eeb81ad0ceda46832e0160/b2_eQV04B8xSdYJZnB2FD.png" width="95%" alt="Light-IF-32B" />
</div>
<hr>
<div align="center">
[](https://arxiv.org/abs/2508.03178)
[](https://huggingface.co/collections/qihoo360/light-if-688753563c174e9d14d6978f)
[](https://zhuanlan.zhihu.com/p/1936535948360918628)
[](https://opensource.org/licenses/Apache-2.0)
</div>
> 本项目用于训练[Light-IF-32B](https://huggingface.co/qihoo360/Light-IF-32B)的冷启动数据可直接参考`cold-start.json`文件获取。
## 🧪 基准测试集
|模型|SuperClue|IFEval|CFBench|IFBench|
| ---- | ---- | ---- | ---- | ---- |
|通义千问3-4B（Qwen3-4B）|0.225|0.888|0.787|0.382|
|通义千问3-8B（Qwen3-8B）|0.225|0.888|0.813|0.417|
|通义千问3-32B（Qwen3-32B）|0.234|0.877|0.823|0.384|
|通义千问3-235B-A22B（Qwen3-235B-A22B）|0.244|0.882|0.834|0.423|
|通义千问3-235B-A22B-Thinking-2507（Qwen3-235B-A22B-Thinking-2507）|0.434|0.916|0.843|0.475|
|深度求索R1-0528（DeepSeek-R1-0528）|0.436|0.863|0.827|0.415|
|豆包-seed-1-6-thinking-250615（Doubao-seed-1-6-thinking-250615）|0.362|0.832|0.82|0.477|
|豆包-seed-1-6-thinking-250715（Doubao-seed-1-6-thinking-250715）|0.345|0.856|0.84|0.366|
|ChatGPT-4o-latest|0.260|0.836|0.807|0.365|
|深度求索v3-250324（Deepseek-v3-250324）|0.306|0.859|0.833|0.405|
|豆包1.5-pro-32k-250115（Doubao-1.5-pro-32k-250115）|0.285|0.889|0.797|0.375|
|Kimi-K2|0.227|0.921|0.820|0.395|
|智谱GLM-4.5|0.395|0.893|0.833|0.466|
| [**Light-IF-4B（本文提出）** 🤗](https://huggingface.co/qihoo360/Light-IF-4B) |**0.445**|**0.916**|**0.80**|**0.443**|
| [**Light-IF-8B（本文提出）** 🤗](https://huggingface.co/qihoo360/Light-IF-8B) |**0.471**|**0.932**|**0.82**|**0.473**|
| [**Light-IF-32B（本文提出）** 🤗](https://huggingface.co/qihoo360/Light-IF-32B) |**0.575**|**0.938**|**0.85**|**0.575**|
## 引用
bibtex
@misc{lightifproj,
title={Light-IF: Endowing LLMs with Generalizable Reasoning via Preview and Self-Checking for Complex Instruction Following},
author={Chenyang Wang, Liang Wen, Shousheng Jia, Xiangzheng Zhang, Liang Xu},
year={2025},
eprint={2508.03178},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2508.03178},
}
提供机构：
maas
创建时间：
2025-10-16
