Suhyunlee/

MHSafeEval

📄 Paper: https://arxiv.org/abs/2604.17730

🎉 Accepted at ACL 2026 Findings

Official implementation of "MHSafeEval: Role-Aware Interaction-Level Evaluation of Mental Health Safety in Large Language Models". MHSafeEval is a closed-loop, agent-based framework for evaluating LLM safety in mental health counseling through adversarial multi-turn interactions, guided by a role-aware harm taxonomy.… See the full description on the dataset page: https://huggingface.co/datasets/Suhyunlee/MHSafeEval.