Mental health LLM Chatbot
Developed this project while leading the ML team at Omdena.
An innovative chatbot designed to diagnose mental health issues such as social media addiction, social isolation, and cyberbullying, based on responses to a customized questionnaire.
The solution involved several key technical strategies:
Custom Dataset Creation:
Leveraged the Gemini and Google Bard models to generate a unique dataset simulating various mental health scenarios, facilitating the chatbot’s ability to provide personalized advice and support.
Model Selection and Fine-Tuning:
Chose Mistral Instruct 7B for its robust capabilities, performing supervised fine-tuning with our custom dataset for enhanced specificity to mental health conditions.
Advanced Model Refinement:
Employed Direct Preference Optimization (DPO) within a Reinforcement Learning from Human Feedback (RLHF) framework, further aligning the chatbot’s responses with user needs. Parameter Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA) techniques were utilized for efficient training on a single NVIDIA A100 GPU.
Deployment Strategy:
The chatbot was deployed using Streamlit on a Google Cloud Platform (GCP) Virtual Machine (VM). To ensure global accessibility, firewall rules were modified, allowing users worldwide to interact with the Streamlit application seamlessly.
This comprehensive approach not only optimized the chatbot for accurately identifying and addressing users’ mental health challenges but also demonstrated a model for deploying AI-driven health support tools globally.