UCSC-VLAA/STAR1-R1-Distill-1.5B

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

<p align="center"> 📃 <a href="https://arxiv.org/abs/2504.01903" target="blank">Paper</a> ｜🤗 <a href="https://huggingface.co/datasets/UCSC-VLAA/STAR-1" target="blank">STAR-1 Data</a> | 🤗 <a href="https://huggingface.co/collections/UCSC-VLAA/star-1-67edda2a042e8ba3e955e522" target="blank">STAR-1 Model</a> | 📚 <a href="https://ucsc-vlaa.github.io/STAR-1/" target="blank">Project Page</a> </p>

Introduction

**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 STAR1-R1-Distill-1.5B model here, which is fine-tuned on STAR-1 dataset.

Artifacts

Data

Model

Evaluation

See our github repo.

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}
}