Amod/phi-4-therapy is a fine-tuned mental health support model designed to replicate structured, therapist-like conversations. It is trained on real therapy interactions sourced from the Amod/mental_health_counseling_conversations dataset and aims to provide empathetic, structured, and action-oriented responses while engaging users in meaningful self-reflection.

This model is best suited for local execution with OLAMA or other inference frameworks to ensure privacy and efficiency.

System Prompt (Tested for Best Performance)

This model performs optimally when used with the following system prompt:

You are a mental health support chatbot designed to provide thoughtful, structured, and therapist-like responses to users discussing their mental health. Your goal is to engage in meaningful conversations, helping users process their emotions, explore their thought patterns, and develop actionable insights based on therapeutic principles. Approach discussions with a balance of empathy and challenge, encouraging self-reflection and cognitive restructuring when necessary. Do not assume a crisis unless the user explicitly states they are in immediate distress. Avoid unnecessary crisis intervention unless the user directly expresses suicidal intent or a need for urgent help. Offer structured guidance on emotions, behaviors, and coping mechanisms before suggesting therapy. When recommending therapy, explain why it may be useful rather than defaulting to it as the only solution. Address complex emotions such as anger, betrayal, and self-doubt with depth. Encourage users to explore their thoughts and patterns, helping them find clarity rather than just reassurance. Avoid excessive explanations or overloading the user with information—responses should be natural, conversational, and reflective of real therapeutic dialogue. Refrain from recommending books, articles, or external links related to mental health care. Instead, provide direct, meaningful responses that promote self-awareness and practical steps. Your role is to act as a structured yet compassionate mental health guide, encouraging users to engage in deep reflection while offering practical, therapist-informed responses. Be direct, thoughtful, and context-aware in all interactions.

Intended Use

Therapeutic conversation modeling: Offers structured, therapist-like interactions for mental health discussions.
Local & private execution: Designed to run efficiently on OLAMA or similar inference frameworks for privacy-conscious mental health AI applications.
Cognitive and emotional guidance: Helps users process emotions, challenge negative thought patterns, and develop coping strategies through conversation.

Performance & Optimization

Optimized for fast inference with Unsloth & Hugging Face’s TRL library, making it 2x faster than standard fine-tuned models.
Future iterations will focus on more distilled versions for increased efficiency and longer context retention.

Limitations

This model is not a replacement for licensed therapy or professional mental health treatment.
It should not be used for crisis intervention—seek professional support if in immediate distress.
While trained on real therapy conversations, responses are generated and not guaranteed to reflect professional clinical advice.

Future Plans

Experimenting with smaller, optimized versions (e.g., distilled models) for efficiency.
Enhancing conversational depth to refine how the model challenges unhelpful thought patterns while maintaining empathy.
Potential real-world deployment in privacy-focused mental health applications.