---
base_model: unsloth/phi-4-unsloth-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- gguf
license: creativeml-openrail-m
language:
- en
datasets:
- Amod/mental_health_counseling_conversations
---

# Amod/phi-4-therapy

- **Developed by:** Amod  
- **License:** apache-2.0  
- **Fine-tuned from:** unsloth/phi-4-unsloth-bnb-4bit  

## Overview  
`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.  

---

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)