STAR-benign-915

# 🌟 STAR-1: Safer Alignment of Reasoning LLMs with 1K Data

📃 Paper ｜🤗 STAR-1 Data | 🤗 STAR-1 Model | 📚 Project Page

## Introduction
[**STAR-1**](https://huggingface.co/datasets/UCSC-VLAA/STAR-1) is a high-quality safety dataset designed to enhance safety alignment in large reasoning models (LRMs) like DeepSeek-R1.

- Built on the principles of diversity, deliberative reasoning, and rigorous filtering, STAR-1 integrates and refines data from multiple sources to provide policy-grounded reasoning samples.
- The dataset contains **1,000** carefully selected examples, each aligned with best safety practices through GPT-4o-based evaluation.
- Fine-tuning with STAR-1 leads to significant safety improvements across multiple benchmarks, with minimal impact on reasoning capabilities.

We open-sourced our [STAR-benign-915](https://huggingface.co/datasets/UCSC-VLAA/STAR-benign-915) (STAR-1's benign variants for mitigating overrefusal) here.

## Artifacts
### Data

| Dataset | Num. of Sample | URL |
|------------|----------------|----------------------------------------------------------------------|
| STAR-1 | 1K | 🤗 [UCSC-VLAA/STAR-1](https://huggingface.co/datasets/UCSC-VLAA/STAR-1) |
| STAR 41K | 41K | 🤗 [UCSC-VLAA/STAR-41K](https://huggingface.co/datasets/UCSC-VLAA/STAR-41K) |
| STAR-benign-915 | 915 | 🤗 [UCSC-VLAA/STAR-benign-915](https://huggingface.co/datasets/UCSC-VLAA/STAR-benign-915) |

### Model
| Model | Type | URL |
|--------------------------------|-------------------------------------------|----------------------------------------------------------------------------------------|
| `STAR1`-R1-Distill-1.5B | R1-Distill-Qwen-1.5B trained on STAR-1 | 🤗 [UCSC-VLAA/STAR1-R1-Distill-1.5B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-1.5B) |
| `STAR1`-R1-Distill-7B | R1-Distill-Qwen-7B trained on STAR-1 | 🤗 [UCSC-VLAA/STAR1-R1-Distill-7B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-7B) |
| `STAR1`-R1-Distill-8B | R1-Distill-Llama-8B trained on STAR-1 | 🤗 [UCSC-VLAA/STAR1-R1-Distill-8B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-8B) |
| `STAR1`-R1-Distill-14B | R1-Distill-Qwen-14B trained on STAR-1 | 🤗 [UCSC-VLAA/STAR1-R1-Distill-14B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-14B) |
| `STAR1`-R1-Distill-32B | R1-Distill-Qwen-32B trained on STAR-1 | 🤗 [UCSC-VLAA/STAR1-R1-Distill-32B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-32B) |

## Acknowledgement
This work is partially supported by a gift from Open Philanthropy. We thank the NAIRR Pilot Program and the Microsoft Accelerate Foundation Models Research Program for supporting our computing needs.

## Citation
```
@article{wang2025star1saferalignmentreasoning,
title={STAR-1: Safer Alignment of Reasoning LLMs with 1K Data},
author={Zijun Wang and Haoqin Tu and Yuhan Wang and Juncheng Wu and Jieru Mei and Brian R. Bartoldson and Bhavya Kailkhura and Cihang Xie},
year={2025},
journal = {arXiv preprint arXiv:2504.01903}
}

# 🌟 STAR-1：基于1K数据实现推理型大语言模型的安全对齐

📃 论文 | 🤗 STAR-1 数据集 | 🤗 STAR-1 模型 | 📚 项目主页

## 引言
[**STAR-1**](https://huggingface.co/datasets/UCSC-VLAA/STAR-1) 是一款高质量安全数据集，旨在提升DeepSeek-R1等推理型大语言模型（Large Reasoning Models, LRMs）的安全对齐性能。

- 本数据集遵循多样性、审慎推理与严格筛选的原则，整合并优化多源数据，生成符合安全政策规范的推理样本。
- 该数据集包含1000个精心筛选的样本，所有样本均通过基于GPT-4o的评估，符合最优安全实践标准。
- 使用STAR-1进行微调可在多个基准测试中实现显著的安全性能提升，同时对推理能力的影响极小。

我们已开源用于缓解过度拒绝问题的STAR-1良性变体数据集[STAR-benign-915](https://huggingface.co/datasets/UCSC-VLAA/STAR-benign-915)。

## 研究产物
### 数据集
| 数据集名称 | 样本数量 | 下载链接 |
|------------|----------------|----------------------------------------------------------------------|
| STAR-1 | 1K | 🤗 [UCSC-VLAA/STAR-1](https://huggingface.co/datasets/UCSC-VLAA/STAR-1) |
| STAR 41K | 41K | 🤗 [UCSC-VLAA/STAR-41K](https://huggingface.co/datasets/UCSC-VLAA/STAR-41K) |
| STAR-benign-915 | 915 | 🤗 [UCSC-VLAA/STAR-benign-915](https://huggingface.co/datasets/UCSC-VLAA/STAR-benign-915) |

### 模型
| 模型名称 | 模型类型 | 下载链接 |
|--------------------------------|-------------------------------------------|----------------------------------------------------------------------------------------|
| `STAR1`-R1-Distill-1.5B | 基于STAR-1数据集微调的R1-Distill-Qwen-1.5B | 🤗 [UCSC-VLAA/STAR1-R1-Distill-1.5B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-1.5B) |
| `STAR1`-R1-Distill-7B | 基于STAR-1数据集微调的R1-Distill-Qwen-7B | 🤗 [UCSC-VLAA/STAR1-R1-Distill-7B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-7B) |
| `STAR1`-R1-Distill-8B | 基于STAR-1数据集微调的R1-Distill-Llama-8B | 🤗 [UCSC-VLAA/STAR1-R1-Distill-8B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-8B) |
| `STAR1`-R1-Distill-14B | 基于STAR-1数据集微调的R1-Distill-Qwen-14B | 🤗 [UCSC-VLAA/STAR1-R1-Distill-14B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-14B) |
| `STAR1`-R1-Distill-32B | 基于STAR-1数据集微调的R1-Distill-Qwen-32B | 🤗 [UCSC-VLAA/STAR1-R1-Distill-32B](https://huggingface.co/UCSC-VLAA/STAR1-R1-Distill-32B) |

## 致谢
本研究部分得到开放慈善（Open Philanthropy）捐赠的支持，同时感谢NAIRR试点项目与微软加速基础模型研究计划（Microsoft Accelerate Foundation Models Research Program）为我们提供计算资源支持。

## 引用

@article{wang2025star1saferalignmentreasoning,
title={STAR-1: Safer Alignment of Reasoning LLMs with 1K Data},
author={Zijun Wang and Haoqin Tu and Yuhan Wang and Juncheng Wu and Jieru Mei and Brian R. Bartoldson and Bhavya Kailkhura and Cihang Xie},
year={2025},
journal = {arXiv preprint arXiv:2504.01903}
}