STAR-1 introduces a high-quality safety alignment dataset specifically designed for large reasoning models, achieving 40% safety performance improvement across benchmarks with just 1,000 samples.

• Built on three core principles: diversity, deliberative reasoning, and rigorous filtering
• Integrates existing open-source safety datasets from diverse sources
• Implements a GPT-4o-based safety scoring system to evaluate and enhance model responses
• Demonstrates effective safety alignment with significantly less data than traditional methods

This research addresses critical security concerns in LLMs by providing a practical approach to safety alignment that's both resource-efficient and highly effective, making safer AI deployment more accessible.

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