Language coverage

The supported set is the 24 official languages of the European Union plus Turkish and Azerbaijani. English is already an EU official language, so the total is 26.

What is measured

Every classifier is scored separately in all 26, at its calibrated threshold, on corpora generated per language rather than translated. There is no aggregate number, because an aggregate hides the tail and the tail is the point.

pii's default model, piiguard, is trained on all 26: read the current set off registry.MODELS["piiguard"].trained_languages rather than off this file, since it moves when the pinned weights do.

pii's second, policy-selectable model does not carry that claim. Setting pii: { options: { model: cee-pii } } in a policy switches to flowxai/cee-pii, a GLiNER model trained on Romanian, Polish, Hungarian, Uzbek and both English variants. Four of those, en, ro, pl and hu, are in the 26; Uzbek is not a language this library claims at all. No per-language evaluation exists for cee-pii yet, so it has no row in the tables below: docs/reference/performance.md reports its latency and says quality: not recorded rather than a score. Selecting it also means the deployment needs a GPU, which registry.deployment_notes(policy) reports the moment a policy switches it on.

Where it fails

Maltese is absent from the base model's pretraining, and that is a fact about the base model, not a working explanation for a weak score by itself. This file said Maltese scored zero on two classifiers and that no amount of data would fix it, which was true of the corpus that existed at the time: those scores rested on two positive examples per language, and grew to ten. docs/reference/performance.json is the current, regenerated reading, and Maltese is not near zero on any classifier in it today. Before attributing a weak score to the pretraining gap, check the sample size it rests on.

Maltese and Azerbaijani national identifiers have no public checksum scheme, so those two can only be generated format-valid, which makes their labels weaker than the rest by construction.

What not to claim

The English locale in the training generator is labelled United Kingdom but uses a German identifier algorithm as a generic numeric fallback. A real UK National Insurance number carries no checksum, so a fallback is defensible, but do not state that English national identifiers are checksum validated.