Compound question-answering model, evaluated

The brief

Scientists ask quick questions about compounds all day. Fine-tune a small open model on question and answer pairs built from public compound data, and evaluate it against a large hosted model: accuracy, made-up facts, speed and cost. Good looks like: a training set built from a trusted source, a test set with known answers, and a careful hallucination check.

What you will build

1. Build question and answer pairs from PubChem records
2. Fine-tune a small model with LoRA
3. Test both models on held-out compounds
4. Measure accuracy and made-up facts
5. Recommend one, with its limits

Free, open data to build it with

• PubChem ↗ National Library of Medicine

Check each dataset's licence before you publish your work. Your mentor gives your team its own version of the brief (data, scope or region), so every team's project is its own.