What is SoulChat2.0
SoulChat2.0 is a digital twin large language model for psychological counselors launched by the School of Future Technology - Guangdong Provincial Key Laboratory of Digital Twin Humans at South China University of Technology, based on the SoulChat1.0 model. It defines for the first time the digital twin task of a specific psychological counselor, aiming to improve the application performance of large models in real psychological counseling scenarios by simulating the language style and therapeutic techniques of real psychological counselors. SoulChat2.0 can build a mental health large model with the style of a specific psychological counselor at low cost, quickly, and efficiently, assisting psychological counselors in their work, such as pre-counseling conversations and providing 24-hour online services.
Main Features of SoulChat2.0
- Digital Twin Modeling of Psychological Counselors: SoulChat2.0 can achieve digital twinning of a specific psychological counselor through fine-tuning, simulating the counselor's language style and counseling techniques, and improving the application performance of LLMs in real psychological counseling scenarios.
- Low-cost, Fast, and Efficient Model Construction: SoulChat2.0 can use a small number of real counseling cases for digital twin modeling of psychological counselors, building a mental health large model with the style of a specific psychological counselor at low cost, quickly, and efficiently.
- Assisting Psychological Counseling Work: SoulChat2.0 can assist real-world psychological counselors in their work, such as conducting pre-counseling conversations and providing 24-hour online services.
- Open Source Collaboration: The project method code, dataset, and model of SoulChat2.0 have been fully open-sourced, facilitating comparison and reproduction by the research community and promoting the development of the mental health large model field.
Technical Principles of SoulChat2.0
- Digital Twin Data Generation for Psychological Counselors: SoulChat2.0 generates digital twin data for psychological counselors for single-turn dialogues by integrating the language style, counseling techniques, and client Big Five personality traits of real-world counselors, combined with real-world counseling cases. This process can quickly generate batches of counseling cases used for the digital twin modeling of that psychological counselor.
- Multi-turn Dialogue Synthesis: SoulChat2.0 builds a digital twin of a psychological counselor with a personalized counseling style based on large language models (LLMs). It involves dynamic one-shot learning from counseling cases, client personality simulation, and multi-turn mental health dialogue synthesis.
- Fine-tuning Large Language Models: Given counseling case data for digital twin modeling of a psychological counselor, SoulChat2.0 can achieve digital twinning of that counselor through fine-tuning. This process involves using a specific base model and performing full fine-tuning on the training set of PsyDTCorpus to achieve the digital twin of a psychological counselor with a specific counseling style.
- Multi-dimensional Evaluation and Improvement: SoulChat2.0 shows significant improvements over other models in dimensions such as counseling techniques, state and attitude, emotional empathy, and cognitive empathy, enhancing the model's psychological counseling performance.
Project Address of SoulChat2.0
- Github Repository：https://github.com/scutcyr/SoulChat2.0
Application Scenarios of SoulChat2.0
- Psychological Counseling Assistance: SoulChat2.0 can simulate the language style and therapeutic techniques of real psychological counselors, assisting real-world psychological counselors in their work. It can be used to conduct pre-counseling conversations and provide 24-hour online services, meeting the needs of users at different time periods.
- Mental Health Large Model Research: SoulChat2.0 uses a small number of real consultation cases to build digital twin models of psychological counselors, bringing new research ideas to the field of mental health large models.