Domain-Specific Assistant: Medical AI

This project implements a Medical AI Assistant by fine-tuning the TinyLlama-1.1B model on the Medical Flashcards dataset using LoRA (Low-Rank Adaptation).

Project Structure

• Medical_Assistant_Fine_Tuning.ipynb: Jupyter Notebook for data preprocessing, model fine-tuning, and evaluation. tailored for Google Colab (Free Tier).
• app.py: Gradio web application for interacting with the fine-tuned model.
• requirements.txt: List of Python dependencies.

Setup Instructions

1. Fine-Tuning (Google Colab)

1. Upload Medical_Assistant_Fine_Tuning.ipynb to Google Colab.
2. Change the runtime type to T4 GPU.
3. Run all cells to fine-tune the model.
4. The fine-tuned adapter will be saved. You can download the medical_assistant_adapter folder.

2. Running the Web App (Local)

1. Install dependencies:

   pip3 install -r requirements.txt

2. Ensure you have the medical_assistant_adapter folder in the same directory (or update the path in app.py).
3. Run the app:

   python3 app.py

4. Open the provided local URL in your browser.

Note: If you are running on macOS or a machine without NVIDIA GPU, the app will automatically load the model in full precision (which requires ~4GB RAM) instead of 4-bit mode.

Methodology

• Model: TinyLlama-1.1B-Chat-v1.0 (Quantized to 4-bit)
• Dataset: medalpaca/medical_meadow_medical_flashcards
• Technique: Parameter-Efficient Fine-Tuning (PEFT) using LoRA.
• Metrics: BLEU and ROUGE scores for evaluation.

Performance

The fine-tuned model demonstrates improved ability to answer medical queries compared to the base model, specifically in maintaining the Q&A format and domain relevance.

Medical_Assistant_Fine_Tuning