Field guide

Teaching a 3B model to read receipts, and finding out what that costs

Every part of the system, with the measurement that decided it. Fine-tuning moved field F1 from 0.440 to 0.836 on receipts the model had never seen. Quantizing to 4 bits cost nothing measurable. Two of the most convincing results along the way turned out to be bugs in the pipeline, and both are written up here rather than quietly fixed.