Large Language Models (LLMs) have demonstrated impressive capabilities across complex tasks, yet their application in specialized medical domains, particularly mental health, presents unique challenges. With global mental health concerns on the rise, LLMs hold significant potential to provide support.
Researchers have identified three primary hurdles for LLMs in mental health applications: the scarcity of high-quality, interpretable, and knowledge-grounded training data; training paradigms often restricted to core functionalities; and insufficient evaluation methods for multi-turn dialogue settings.
To address these critical issues, a new "oMind" framework has been introduced. This framework aims to comprehensively enhance LLMs for mental health, featuring several key components:
- **Integrated Training and Alignment:** oMind provides a robust framework for training and aligning LLM agents, equipping them with diverse conversational capabilities.
- **High-Quality SFT Dataset:** The framework includes an extensive ~164k multi-task Supervised Fine-Tuning (SFT) dataset. This dataset is generated through a sophisticated pipeline involving structured knowledge retrieval, LLM-based pruning, and expert review, ensuring both quality and knowledge grounding.
- **oMind-Chat Multi-Turn Benchmark:** oMind-Chat is a novel multi-turn dialogue benchmark dataset. Its distinguishing feature is the inclusion of expert-annotated turn-level and conversation-level rubrics, filling a crucial gap in current evaluation methodologies.
Diverse experiments across both core capabilities and conversational tasks show that oMind LLMs consistently outperform baseline models. Furthermore, oMind-LLM demonstrates significantly improved reasoning abilities, achieving up to an 80% win rate, thus establishing a strong foundation and new direction for LLM development in the mental health sector.