SFT for Mental Health Question-Answering

Fine-tuned Llama2-7b for chat and question-answering on mental health datasets as part of research at the Kelley Data Science and AI Lab, Indiana University.

Approach:

• Supervised fine-tuning using Hugging Face Transformers with QLoRA for parameter-efficient training
• Deployed models using vLLM for efficient inference
• Built a Streamlit-based evaluation platform for end-user interaction, feedback collection, and pairwise comparison evaluation across model responses
• Platform served 50+ researchers with automated ELO-based model ranking

Tech: Python, Transformers, QLoRA, vLLM, Streamlit, SQLite