The encoder-only transformer model — architecture, BIO labeling, decoding, and inference.
View All TagsOur resolver kept dropping German addresses in New Hampshire, and our scorecard handed it a gold star every time. This post shows how a name-match metric lied to us for months, and how much a postcode hint improved placement once we measured by distance instead.
We injected a postcode's country signal straight into the model, and German addresses jumped thirty-five points. Then three retrains taught us the same anchor can't help an address written in the other order, because it never learned to find the city. It learned which way to look.
We thought a postcode was a strong enough signal to delete a configuration flag. Then we measured it. A five-digit code pins its own country less than half the time — and that number talked us out of writing the code we were about to write.
A neural address parser that scores 20% on our test suite and beats Pelias on real addresses. Both numbers are true — here's why we trust the second one.
We set out to fix our address parser's overconfidence. We discovered overconfidence was never the problem — and that you can't calibrate your way out of a coverage gap.
We pulled Who's On First data for 7 countries, trained a new tokenizer on 2.19M multi-script place names, and eliminated CJK byte-fallback entirely.
Japanese addresses run backwards: prefecture, then city, then block, then number. They have no street names. Wepulled 6,373 of them from Who's On First.
v0.6.0 expands the schema from 10 to 16 component tags. PO box recognition: 0% → 51.9% in one training run.
The v0.5.3 model now runs with a 9 MB finite-state transducer in the browser — 94K US admin places as an emission prior for the neural classifier.
A one-line import change fixed a WebGPU inference bug that took hours to diagnose. The root cause: onnxruntime-web ships two WebGPU backends, and the default one is broken.