BluzenTagger

Tagging illustrations is a long-tail problem wearing a classification problem's clothes. The head of the distribution is easy and boring: a model that only ever predicts the two hundred most common tags will score well on any average you compute over images, and be useless to anyone searching for something specific.

What we changed

Three things, none of them clever on their own.

1. Stratified evaluation first. We split the tag vocabulary into frequency bands and report every metric per band. A number that improves on the head while collapsing the tail now looks like what it is.
2. Per-tag thresholds. A single global confidence cutoff is a compromise between thousands of differently-calibrated heads. Calibrating each one against a held-out slice costs nothing at inference.
3. Loss reweighting on band, not on frequency. Straight inverse-frequency weighting made rare-tag training unstable. Weighting by band kept the gradients sane.

from bluzen_tagger import Tagger

tagger = Tagger.from_pretrained("bluzen/bluzen-tagger-v1")
tags = tagger("illustration.png", threshold="calibrated")

Status

Active. v1 is stable and the v2 line is training against a larger Aurora slice.