Finding and Reactivating Post-Trained LLMs' Hidden Safety Mechanisms

Develop a tool to audit and recover safety alignment mechanisms that are often 'bleached' during heavy CoT fine-tuning.

Suggested repo: safe-react

"Find and rescue the safety mechanisms lost in your fine-tuning run."

Estimated effort: 50h