Training runs, loss recipes, verdict smokes, and divergence diagnostics.
View All TagsI taught the address parser three new countries overnight. Two of them landed. The third failed a regression check I trusted, and the reason was concerning: the new addresses never included their own country, and the model was learning to stop labeling countries at all.
We're trained to be suspicious of a score that looks too good. The scores that say you failed get a free pass — and this week three of them were lying, including one that had us reporting kilometers for a product that lands in meters.
We taught our parser the hardest street cases it gets wrong, the aggregate score climbed, and we almost shipped it. Then a held-out probe caught it forgetting what a town is.
Our failure taxonomy now ranks boundary instability first among the parser's weaknesses. We
Two days ago we retired a piece of decode-time implementation we'd been leaning on,
Yesterday we wrote about a lookup table that scored a perfect 100 and nearly talked us
This morning we published a post that ended with a tidy rule: some address tags don't
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 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.