Edge spending categoriser
A 0.5B-parameter language model, fine-tuned on a MacBook Air, that turns messy bank-statement lines into clean JSON, small enough to run on the device so financial data never leaves it.
What it does
Real outputs from the best model on merchants and line formats it never saw in training. Right and wrong.
Results
The gap between seen and unseen merchants is the difference between memorising and generalising. Hover for 95% confidence intervals and significance against the best run.
Explore every prediction
Every test line, what it should be, and what the model said. Click a row to compare all runs on that line.
| Statement line | Expected | Model said | Cat | Merch | Amt |
|---|
Where categories go wrong
True category (rows) against predicted (columns) for the selected run. Click a cell to see those lines.