CEE-PII-Bench (v0.2)

A held-out, multilingual (RO / PL / HU / UZ / EN-UK / EN-US) benchmark for span-level PII / sensitive-entity detection, weighted toward Central & Eastern European identifiers. CEE-PII-Bench is the evaluation companion to the flowxai/cee-pii detector. Its headline metric is false-positive rate on no-entity documents — because a privacy guard that fires on clean text gets disabled within a week, and a disabled guard protects nobody.

Why this benchmark exists

English-centric PII benchmarks under-represent CEE national IDs (CNP, PESEL, TAJ, PINFL), CEE address grammar (the RO Str./nr./bl./sc./et./ap. shape), and the diacritic-optional reality of Romanian text. CEE-PII-Bench evaluates all 34 taxonomy types on the same model-agnostic span scorer used for the detector, the zero-shot GLiNER baseline, and the frontier reference — so the numbers are directly comparable across models.

Composition (v0.2)

Frozen from the full split-aware TEST split of corpus v0.2. Two subsets:

Subset	file	n docs	no-entity docs	definition
standard	v0.2/ceepii_bench_v0.2.jsonl	2,000	274	the whole frozen v0.2 TEST split
hard	v0.2/ceepii_bench_v0.2_hard.jsonl	960	134	has_noise AND has_negatives

Per-language (standard subset)

Language	docs
en_uk	410
pl	385
uz	352
ro	332
en_us	270
hu	251

Balanced across 7 registers (bank statement, chat, contract clause, email, form field, OCR-ish fragment, support ticket).

Entity taxonomy coverage (standard subset)

The taxonomy has 34 types across three tiers (checksum-validated national IDs, structured identifiers, contextual entities — see the detector card for the full list). v0.2 standard exercises 23 of the 34 types with non-zero gold; the remaining 11 (e.g. ein, itin, nino, pesel, ssn, szemelyi, uk_account_number) have zero gold in v0.2 and are a coverage artifact, not a model failure — they need bench coverage in a future freeze before their per-type F1 is meaningful.

Gold-count highlights (standard subset): address 372, account_ref 359, plate 255, person_name 193, policy_ref 181, dob 177, taj 172, email 154, aba 119, employer 100, utr 70, ci_ro 69, postal 66, uz_account 62, nhs 61, pinfl 57, contract_ref 56, phone 53, card 49, cnp 38, first_name 12, health_condition 9, surname 5.

Record schema (JSONL)

Each line is a document with its gold entity spans as (type, start, end) character offsets, plus language/register metadata and the flags used to define the hard subset (has_noise, has_negatives, whether the doc is zero-entity).

How it was built

• Fully synthetic. Generated by the cee-pii-phase3-v0.2 pipeline (corpus seed 20260702) from carrier templates + checksum-valid entity generators, then noised at assembly (diacritic stripping — critical for RO — OCR confusions, casing, whitespace/punctuation damage) with character-level span tracking through every transformation, so gold offsets stay correct after noise.
• Family- AND value-disjoint splits, zero straddlers. Template families are partitioned into train/val/test first (stratified by (language, register), family-split seed 20260703, 80/10/10); each split is then generated from only its own family pool, so straddler_count = 0 by construction. Value-disjointness is enforced by resample with train > val > test priority.
• Contamination-verified. Bench families ∩ train = ∅ and bench values ∩ train = ∅, both asserted by automated tests. v0.2 carries 42 bench-owned families and 2,678 distinct bench values, none of which appear in any training config.
• Manifest (v0.2/MANIFEST.json): per-file sha256, per-language / per-register / per-entity-type stratification, split policy + seeds.
  • ceepii_bench_v0.2.jsonl sha256 15b49cf42411d990660d59bee6769a5d5f449746b31f12a79ed9e226ed214acd (2,000 docs).
  • ceepii_bench_v0.2_hard.jsonl sha256 352e4d3772cecb4e566a7ae808ea9f295a0554c857a7d726dc5f700c7cc6e3c2 (960 docs).

Sanitization & privacy statement

CEE-PII-Bench is a public artifact and contains zero real PII:

• All values are synthetic. Checksum identifiers (CNP, PESEL, IBAN, NHS, ABA, …) are generated to be internally valid but correspond to no real person. Structured identifiers follow format/range rules only.
• Names are formed by independent sampling of public census / statistical first-name and surname frequency lists — a full-name pair is never copied from any source, so no real individual's name is reproduced.
• No client data, no scraped personal data, anywhere in the benchmark or the repo (including tests).

Honesty caveat (read before citing scores)

CEE-PII-Bench v0.2 is held-out and contamination-verified, but it is drawn from the same synthetic generator distribution as the training corpus. Scores on it are a valid relative signal (comparing models on identical inputs) but are likely optimistic in absolute terms versus real documents — the same home-field-advantage caveat as the sibling scam-guard project. The honest real-world number is a human-labeled real-structure eval set (official form specimens, published template contracts, sample bank statements), reported separately and still to be built.

Versioning

CEE-PII-Bench is versioned and never regenerated silently (v0.1 → v0.2 froze the full split-aware TEST split, ~2,000 docs). Cite the version and manifest sha256 you evaluated against.

License

Apache-2.0.