Roleplaying with Structure: Synthetic Therapist-Client Conversation Generation from Questionnaires

Large Language Models (LLMs) are promising tools for synthetic data generation in mental health, yet previous work is mostly limited to generic seed information because of privacy regulations. We present SQPsych (Structured Questionnaire-based Psychotherapy), a pipeline for generating synthetic therapist-client conversations. SQPsych uses real structured client profiles and psychological questionnaires to generate synthetic corpora, SQPsychConv. We then fine-tune seven open-weight LLMs and evaluate them using automatic benchmarks and trained psychotherapists. Our fine-tuned models remain competitive with baselines on surface-level counseling benchmarks and almost always outperform the previous mental-health state of the art. Expert evaluation further shows that SQPsych significantly improves LLMs’ ability to roleplay therapists, and experts consistently prefer therapy sessions generated by our models over those from other mental-health-oriented LLMs. We release our code, fine-tuned SQPsychLLM models, and synthetic corpora at https://ai-mh.github.io/SQPsych.html.