pl-pii-bench

Maintained by Anonimator.pl, a local-first Polish document anonymization tool. GitHub repository · Live benchmark results

pl-pii-bench is an open benchmark for Polish personally identifiable information detection and text anonymization. It contains a fully synthetic, exhaustively annotated, document-level corpus and a separate open scoring harness.

The corpus is an evaluation set, not training data. Please do not train on it. Every identifier is synthetic. Checksum-bearing Polish identifiers are valid for their formats but do not belong to real people.

It is sized for depth rather than volume, and the item count should be read in that light. The core documents are complete Polish administrative documents, exhaustively annotated for all seventeen labels, not sentence-length samples. Exhaustive annotation is what makes precision valid on every split; a larger corpus annotated only for planted entities reports recall honestly and precision meaninglessly.

Why this benchmark is different

1. Checksum-valid identifiers paired with invalid lookalikes. PESEL, NIP, REGON, dowod osobisty, and IBAN/NRB values pass their real check-digit algorithms, and the negative split carries same-shape values that fail their check digit. Checksum-valid generation alone is not novel; the matched invalid set is what separates a detector that validates from one that only pattern-matches, and it is what makes identifier precision measurable rather than assumed.
2. Polish morphology, reported per case. The inflection split exercises person and organization names across all seven grammatical cases, and recall is stratified by case rather than averaged over them. Other benchmarks annotate inflected names; this one reports where inflection breaks a detector.
3. Polish document types and an extraction-noise split. The corpus covers complete synthetic Polish administrative documents. The PDF split includes both source PDFs and canonical extracted text with character-offset annotations, so ligature glitches, page-boundary-split identifiers, and table reordering are captured in the committed text. Detection is scored against that committed text, identical for every system, not against each system's own extraction.

Splits and corpus size

Lane	Documents
core	33
inflection	441
identifiers	12
address	14
negative	14
robustness	231
pdf	14

This table covers only the public corpus. Its document counts come from the authoritative release manifest. The private holdout corpus is not described here: its per-lane sizes are not part of the public release.

The public address lane has 164 annotated spans, so one missed span changes its span recall by about 0.6 percentage point. Every identifier label carries at least 30 annotated spans across at least four rendered formats.

The seven public splits are core, inflection, identifiers, address, negative, robustness, and pdf. The pdf directory keeps each source PDF next to the JSON document containing the canonical extracted text used for scoring.

Labels

The seventeen labels frozen for v1.0 are PERSON, ORG, PESEL, NIP, REGON, DOWOD, IBAN, LOC, POSTAL, DOB, PHONE, EMAIL, PASSPORT, DRIVING_LICENSE, PAYMENT_CARD, VIN, and VEHICLE_PLATE.

See annotation-guidelines.md for the complete span-boundary, protection, identifier-class, address, and morphology contracts.

Annotation and quality assurance

Annotations are emitted by the generator that produces each document, then validated mechanically. The full corpus is re-validated at the start of every release run, so no engine is scored against an unvalidated corpus.

Enforced as hard failures:

• Offset contract. text[start:end] must equal the entity text exactly for every entity in every document. The scoring harness re-asserts this independently when it loads the corpus, so a drifted offset fails the run instead of silently degrading a score.
• Checksum validity of the ground truth. Every PESEL, NIP, REGON, and DOWOD span in the positive splits is re-derived through its check-digit algorithm. Exemptions are explicit and narrow: robustness fixtures that corrupt check digits by design, each keyed to its clean twin.
• Per-entity metadata. identifier_class must be direct or quasi and protection must be protect or keep, so re-identification metrics cannot be computed over partially tagged documents.

Exhaustiveness, which precision depends on, is checked by reverse error analysis: the open baseline detectors are run across the corpus and every high-confidence prediction that is not in the ground truth is surfaced for manual adjudication. This is the Presidio-research error-analysis loop inverted into a completeness check. It is advisory rather than blocking, because a baseline false positive and a real annotation gap look identical until a human reads them.

There is no inter-annotator agreement figure, and that is deliberate rather than an omission. This corpus is programmatically generated rather than independently hand-annotated, so its failure mode is generator error, not annotator disagreement; a pairwise F1 computed over one generator's own output would be 1.0 by construction and would carry no information. Human-annotated benchmarks do report agreement, and should. Here the correctness guarantee is mechanical and total on the offset, checksum, and metadata contracts, and human only in the adjudication loop above.

Metrics

The open harness reports strict and relaxed span matching per split and label, recall, precision, F2, protected recall, direct-identifier recall, quasi-identifier coverage, residual-identifiability rate, morphology case recall, format recall, robustness deltas, and negative-split false positives as both a raw count and FP per 1000 tokens. Token counts use len(text.split()). There is no single blended aggregate.

Read the per-label table, not an average across labels. Of the seventeen labels, only PERSON, ORG, and LOC genuinely require learned extraction. The other fourteen have deterministic formats, and eight of those carry check digits, so a rule set alone scores highly on them. Any average across all labels is therefore dominated by rule coverage rather than model quality, and two systems with very different extraction ability can land on the same aggregate. This is also why the negative split matters: on deterministic labels it is the only place where validating and pattern-matching separate.

Reproduced v1.0.0 results

The table below is generated from the authoritative local release manifest. Only public-split summaries are copied into this package. Commands, local paths, private evaluation material, predictions, and detailed reports are not included.

The public matrix contains exactly five local engines: Anonimator, Presidio, spaCy PL, GLiNER PII Polish, and BardsAI EU PII. Remote API and terminal-agent systems are not public rows. An optional private Codex subscription diagnostic may inspect public synthetic documents, but it has no score in this card, the public manifest, release packages, or the website.

Split	Documents	Anonimator relaxed recall	Presidio relaxed recall	spaCy PL relaxed recall	GLiNER PII Polish relaxed recall	BardsAI EU PII relaxed recall
core	33	99.5%	95.8%	64.2%	58.0%	69.7%
inflection	441	100.0%	96.4%	96.4%	97.3%	37.0%
identifiers	12	75.0%	44.4%	3.2%	16.7%	31.2%
address	14	93.9%	79.3%	79.3%	29.9%	68.3%
negative*	14	32 FP	362 FP	346 FP	123 FP	513 FP
robustness	231	76.6%	68.0%	56.5%	54.9%	61.6%
pdf	14	99.5%	95.2%	63.1%	55.1%	64.7%

* The negative split has no planted entities, so relaxed recall is undefined; the value shown is the false-positive count instead.

Every row above comes from one pinned tool version. Model-based engines are pinned to an immutable revision, not a mutable branch, so the row is reproducible.

Engine	Reproduction versions
Anonimator	version 0.1.0
Presidio	presidio-analyzer 2.2.364, spaCy 3.8.14, pl_core_news_lg 3.8.0
spaCy PL	spaCy 3.8.14, pl_core_news_lg 3.8.0
GLiNER PII Polish	gliner n/a
BardsAI EU PII	transformers n/a, torch n/a, model bardsai/eu-pii-anonimization-multilang-v2-preview @ 8e0b19766bb0dd4916d096b4f540dd46c138c760

Source commit: 29b51c335b26454b5c9271df9eeb343dcd28a8d4

Public artifact-set SHA-256: 62cedba703ee5197194a3d5db49a8114ea7875d0504060808135e85a00ab20a6

Licenses

The data in data/ and the annotation guidelines are CC BY 4.0. See DATA_LICENSE. The separate scoring harness and baseline adapters are Apache-2.0.

Citation

@misc{pl-pii-bench,
  title  = {pl-pii-bench: An Open Benchmark for Polish PII Detection and Anonymization},
  author = {Anonimator.pl},
  year   = {2026},
  note   = {Version 1.0.0},
  url    = {https://huggingface.co/datasets/pl-pii-bench/pl-pii-bench}
}

Links

• Maintainer: Anonimator.pl
• Source code and open harness: github.com/pl-pii-bench/pl-pii-bench
• Live, human-readable results: anonimator.pl/benchmark.html

Related locale-specific PII benchmarks

This benchmark occupies the Polish slot in a small family of locale-specific PII evaluation sets. The comparison is offered so the differences are legible rather than implied.

Benchmark	Locale	Items	Text length	Negative material	Per-case morphology	Private holdout
pl-pii-bench	Polish	755 public	full documents	dedicated split, invalid checksums	yes	yes
pii-bench-zh	Chinese	8,000	sentences, chat messages	none	n/a	no
tw-PII-bench	Traditional Chinese	910	15–5,000 chars	hard negatives, 40 items	n/a	no
hivetrace/pii-bench	Russian	1,810	6–520 chars	42% of the scenario split	no	no

Two properties here are not present in any of the others: recall stratified by grammatical case, and a private holdout twin generated from the same axes with a different seed, which turns "please do not train on this" from a request into a measurement. Two properties of the others are worth knowing when reading these numbers: tw-PII-bench covers longer texts than its item count suggests, and hivetrace/pii-bench is human-annotated with a published agreement figure, which this corpus does not have.

Known limitations

• The private holdout address material is distinct from the public lane.
• The residual-identifiability rules are fixed deterministic combinations, not a probabilistic linkage adversary.
• End-to-end PDF extraction is not part of the comparison. The PDF split measures detection over one committed canonical extraction.

Feedback

The maintainer of this benchmark also maintains one of the systems it measures (see the harness repository's "Known biases" section for where that shows up). Outside scrutiny is how a benchmark like that earns trust rather than asserts it. Open an issue on github.com/pl-pii-bench/pl-pii-bench for anything that would make a split more thorough or more objective: a labeling rule you think is wrong, an annotation gap, a matching or scoring edge case, or a split you think is missing.