STAIR: Improving Safety Alignment with Introspective Reasoning

🚧 Codebase, datasets & models are coming soon!
⭐ Star to stay updated on our release progress!

Official implementation of STAIR, the framework presented in our paper "Improving Safety Alignment with Introspective Reasoning". STAIR enhances LLM safety with the incorporation step-by-step analysis of potential risks, providing more robust alignment while better maintaining model capabilities.

TODO List

• Official implementation of STAIR, including Safety-Informed MCTS (SI-MCTS), test-time scaling, etc.
• 🤗 SFT dataset of structured CoT format alignment
• 🤗 Model weights of LLMs (Llama-3.1-8B-Instruct, Qwen2-7B-Instruct) aligned with STAIR