PRJ-06REV 2026.01
Condition-keyword prediction LLM
A fine-tuning study. Llama 3.2 3B, trained with QLoRA on 100K synthetic examples, predicts a condition keyword from symptoms.
- tags
- AILive
- stack
- Llama 3.2 3B · QLoRA · Synthetic data · Google Colab · Weights & Biases · Hugging Face · Modal · Gradio
Problem
Can a small open-source LLM learn a narrow, well-defined task end to end, from dataset to fine-tune to a live endpoint? I picked a structured task: given free-text symptoms plus age group, severity and duration, predict the most relevant condition keyword.
What I built
- Generated a synthetic dataset of 100K examples.
- Fine-tuned Llama 3.2 3B with QLoRA on Google Colab, tracking runs in Weights & Biases.
- Published the model to Hugging Face and deployed it on Modal, with a Gradio interface and API.
The repo has the data-analysis notebook and setup instructions.
Trained on synthetic data. It’s a demonstration of the fine-tuning workflow, not for clinical use. The demo starts on demand, so the first request can take a minute or two.