What AI says about Safety-J

Safety-J is a bilingual safety evaluator that assesses the safety of content produced by large language models, delivering detailed critiques and safety judgments. It operates in English and Chinese, uses critique-based evaluation, and employs iterative preference learning to improve over time across a wide range of safety scenarios. The project includes an automated benchmark and leaderboard to evaluate critique quality and compare model safety performance, and claims state-of-the-art results compared with open-source models and GPT-4o on safety tasks.