PRJ-06REV 2026.01
Medical keyword extraction LLM
Llama 3.2 3B fine-tuned with QLoRA on 100K synthetic examples to pull medical keywords from symptom descriptions.
- tags
- AILive
- stack
- Llama 3.2 3B · QLoRA · Synthetic data · Modal · Gradio
Problem
Can a small open-source model be adapted to a narrow, well-defined task through fine-tuning? Extracting medical keywords from symptoms and patient information is a good test. The goal isn’t diagnosis, just structured extraction.
What I built
- Created a synthetic dataset of 100K examples.
- Fine-tuned Llama 3.2 3B with QLoRA for the extraction task.
- Deployed the fine-tuned model on Modal, with a Gradio interface to try it.
The repo has the project write-up and analysis.