EuroPriv-Bench (v0 — detection)

Held-out gold for pan-European PII/PHI de-identification, labeled in the harmonized KP taxonomy (BIOES). Each row: text, spans ({start, end, label}, KP labels), language. Scored by the europriv-bench harness (entity F1 / recall-weighted F2; re-identification-risk + privacy-utility tracks to follow).

v0 scope: seeded from AI4Privacy open core, remapped to KP. Legal + clinical splits, Romanian (not in AI4Privacy), and TAB/MEDDOCAN subsumption land in later phases. This is the first unified step, not the finished benchmark.

Configs

One per language: en, fr, es, de, it, nl (split: test).

from datasets import load_dataset
ds = load_dataset("klusai/europriv-bench", "en", split="test")

Attribution & license

Derived from AI4Privacy open-pii-masking-500k (CC-BY-4.0); native labels remapped to the KP taxonomy crosswalk. Redistributed under CC-BY-4.0 with attribution. Only cleanly-licensed sources are used so the suite stays openly redistributable.

Romanian config: ro-realskeleton-v1

The citable Romanian track: documents that mirror the STRUCTURE and boilerplate of real RO official document types — the CNAS SCRISOARE MEDICALĂ discharge letter, a services contract, a DECLARAȚIE PE PROPRIA RĂSPUNDERE, and an administrative letter — populated with synthetic RO identifiers (valid-checksum CNP, CNP-consistent date of birth, RO IBAN/CUI/CI, county addresses, +40 phones), with Romanian gender agreement. Because the skeletons are authored faithful reproductions of public document structure (RO official texts are non-copyright; Law 8/1996 art. 9(b)) and all identifiers are synthetic, the artifact carries no real personal data and is CC-BY-redistributable.

Pair with ro-synthetic-v1 (toy-template development track): the F1 / CNP-leakage gap between the two is the synthetic-context vs real-context measurement (Paper 2). Scored with cnp_leakage (a missed CNP deterministically discloses DOB + sex + county).

Polish config: pl-realskeleton-v1

A citable-track candidate (currently config_status=dev; not yet validated — pending native-speaker review + IAA, KLU-27) and the second, zero-shot decode-bearing measurement (after ro-realskeleton-v1): documents that mirror the STRUCTURE and boilerplate of real Polish official document types — the hospital discharge card KARTA INFORMACYJNA LECZENIA SZPITALNEGO, a services contract UMOWA O ŚWIADCZENIE USŁUG, a declaration OŚWIADCZENIE, and an administrative letter — populated with synthetic Polish identifiers (valid-checksum PESEL, PESEL-consistent date of birth, NIP/REGON/IBAN, dowód osobisty, addresses, +48 phones), with Polish gender agreement. Because the skeletons are authored faithful reproductions of public document structure (Polish official documents/materials — dokumenty i materiały urzędowe — are excluded from copyright under Art. 4 of the Polish Copyright Act, ustawa o prawie autorskim i prawach pokrewnych; the document structure is therefore non-copyright) and all identifiers are synthetic, the artifact carries no real personal data and is CC-BY-redistributable.

Every row carries country="PL" so the country-dispatched national_id_leakage metric validates and decodes the gold IDs with the PESEL validator (a missed PESEL deterministically discloses DATE_OF_BIRTH + SEX). The protection result here is zero-shot: kp-deid-mdeberta-280m leaks 0% PESEL on PL despite being RO-trained and never having seen Polish (zero-shot transfer, not PL-trained performance). The protection-vs-leak contrast on PL is carried by GLiNER (~58%) and tabularisai (~31%); note openai/privacy-filter also leaks ~0% PESEL on PL, so the "type-accurate detectors leak" story holds for 2 of 4 baselines here, not universally. All four templates are one authored skeleton family sharing a single fill path — a leak headline from a single template family is not validated generalization, so a second independent template family is required before this is cited. The discharge card legitimately repeats the patient PESEL (identity header + Identyfikator pacjenta); re-identification is counted per distinct subject, so the repeat collapses to one subject (KLU-49 dedup) and never double-counts.

Pair with ro-realskeleton-v1 (the first decode-bearing track): together they test whether the train-for-protection result holds on a SECOND identifier (PESEL) in a SECOND language (PL).