TonyYun/pii-shield-benchmark

PII Shield Benchmark

Five public PII-detection datasets rewritten into one record shape, so models can be trained and evaluated across all of them without writing five parsers. 2,301,892 records carrying 13,343,257 labeled spans in 25 European languages.

Nothing here is new data. Every text and every annotation comes from one of the five source datasets below; this repository only harmonizes the containers: one schema, one label taxonomy, one file format (Parquet, zstd). The original labels ride along on every span, so any harmonization decision can be reversed.

Record shape

Every config yields records with the same fields:

{
  "uid": "nemotron-pii/train/000042",
  "source": "nemotron-pii",
  "native_split": "train",
  "language": "en",
  "text": "... the raw document text ...",
  "spans": [
    {
      "entity": "PHONE_NUMBER",
      "sub": "fax",
      "start": 118,
      "end": 132,
      "text": "+977 1 4582 3941",
      "native_label": "fax_number",
      "flags": []
    }
  ],
  "meta": { "domain": "healthcare", "locale": "us" }
}

• start / end are Unicode code points into text, end-exclusive. The invariant text[start:end] == span.text holds for every span in the dataset and is verified at build time.
• entity is one of 47 shared entity types (list below). sub optionally carries a finer role (for example given vs family under PERSON_NAME), and native_label is always the source dataset's own label.
• native_split is the source's original split name (train/validation/test/dev/all), preserved as data. All records of a config are published under a single train split; carve your own subsets from native_split or uid.
• meta passes source-specific extras through (domain, locale, document type, template id, and so on); its fields differ per config.
• flags marks known data-quality caveats, explained below. An empty list means none.

Sources

config	source dataset	license	contents
openpii-1m	ai4privacy/pii-masking-openpii-1m	CC-BY-4.0	1,428,143 synthetic records, 23 languages, 10,328,208 spans
nemotron-pii	nvidia/Nemotron-PII	CC-BY-4.0	200,000 synthetic business documents, English, 1,675,796 spans
gretel-pii-en-v1	gretelai/gretel-pii-masking-en-v1	Apache-2.0	60,000 synthetic documents, English, 254,706 spans
privy	piimb/privy	MIT	602,869 synthetic protocol payloads (large config), English, 1,069,918 spans
mapa-sentences	piimb/mapa-eur-lex-pii	CC-BY-4.0	10,838 sentences from EUR-LEX court judgments, 21 languages, 6,142 spans
mapa-testdocs	piimb/mapa-eur-lex-pii	CC-BY-4.0	42 full EUR-LEX judgments with a finer 13-label taxonomy, 8,487 spans

The two mapa configs are the only human-written text in the pool; everything else is synthetic. mapa ships two incompatible taxonomies at two granularities, which is why it becomes two configs. 7,564 of the mapa sentences contain no PII at all and are kept as negatives for false-positive measurement.

How it was derived

1. Structural examination. Each source stores annotations differently: character spans in JSON lines, python-repr strings inside Parquet, surface strings without any positions, sentence-level and document-level offsets. Each got a dedicated converter.
2. Label unification. The five sources define 163 distinct labels between them. Each was mapped, with per-label evidence from sampled values, onto 47 shared entity types. Merges never destroy information: the finer distinction moves into sub (for example fax_number becomes PHONE_NUMBER with sub: "fax") and native_label keeps the original. Three privy classes were dropped as non-PII payload: O (2,512,258 spans of filler), FINANCIAL (a mixed bag of currency codes and routing-shaped numbers) and CURRENCY (zero spans).
3. Verification before build. Every record of every source file was scanned: every label known, every offset inside bounds, every stored surface equal to the text slice. The scan established, among other things, that all sources use code-point offsets and that every gretel surface occurs exactly once in its text.
4. Conversion with loud failure. The build refuses to continue on an unknown label or an unexplained text mismatch, and the finished output is read back and checked against span counts pinned from the verification scan, so a converter bug cannot pass silently.

What was deliberately not carried over: the sources' masked-text renditions and token-level BIO/BILOU layers (both regenerable from the spans), and privy's small config (the same generator as its large config, re-sampled with abbreviated label names).

The 47 entity types

group	entities
person	PERSON_NAME HONORIFIC USERNAME GENDER AGE OCCUPATION EDUCATION_LEVEL EMPLOYMENT_STATUS
contact and location	EMAIL_ADDRESS PHONE_NUMBER URL STREET_ADDRESS CITY REGION COUNTRY POSTAL_CODE COORDINATES LOCATION
time	DATE_TIME
government identifiers	NATIONAL_ID TAX_ID PASSPORT_NUMBER DRIVER_LICENSE_NUMBER CERTIFICATE_LICENSE_NUMBER
financial	PAYMENT_CARD_NUMBER CVV PIN BANK_ACCOUNT_NUMBER IBAN BANK_ROUTING_NUMBER SWIFT_BIC MONEY_AMOUNT
network, device, vehicle	IP_ADDRESS MAC_ADDRESS DEVICE_ID VIN LICENSE_PLATE
credentials	PASSWORD API_KEY HTTP_COOKIE
health and special category	HEALTH_ID BLOOD_TYPE BIOMETRIC_REF SPECIAL_CATEGORY
organization and other	ORGANIZATION INTERNAL_ID LANGUAGE

LOCATION is the coarse geographic class (privy's LOCATION, mapa's country-level ADDRESS); the six fine location entities can be rolled up into it for coarse scoring.

Flags

Flags are quality annotations added during conversion; no source dataset carries such a field. They let an evaluation include or exclude known dirt instead of silently absorbing it. Counts over the whole dataset:

flag	spans	meaning
synthesized_offset	254,706	gretel stores no positions; offsets were computed by exact substring search. Verified unambiguous: every surface occurs exactly once in its text.
noisy_source_label	84,835	every privy TITLE span. The label is systematically polluted: 60.3% genuine honorifics, 19.9% sex or gender terms, 19.9% occupations.
boundary_whitespace	30,633	the span begins or ends with whitespace in the source annotation; relevant for strict-boundary scoring.
value_case_mismatch	12,965	nemotron's stored span text differed from the document in letter case only; the document slice wins.
format_invalid	3,795	nemotron SWIFT/BIC values violating ISO 9362 lengths, and vehicle identifiers that are not valid 17-character VINs.
vague_value	2,265	nemotron time spans containing no digits ("office hours").
non_literal_value	563	nemotron email spans containing no @ — the generator annotated references ("via email", "the provided email address"), form placeholders ("[Email Address]"), usernames, and one phone number as email addresses. Provably not literal addresses; exclude for detector scoring.
annotation_bug	65	nemotron vehicle-identifier spans whose text is the literal string license_plate.
offset_ambiguous	0	reserved: would mark a synthesized offset whose surface occurs more than once. Never fires on this data.

Known limitations

• All IBANs in the dataset (16,897 spans, all from privy) are GB-prefixed. There are no EU-format IBAN positives.
• The synthetic sources inherit their generators' quirks; the flags above mark the ones found so far, without any claim of completeness.
• Rule-of-thumb quality differs by source: mapa is human-annotated real text, everything else is template- or LLM-generated.

Usage

from datasets import load_dataset

ds = load_dataset("TonyYun/pii-shield-benchmark", "openpii-1m", split="train")
r = ds[0]
for s in r["spans"]:
    print(s["entity"], repr(r["text"][s["start"]:s["end"]]))

Each config directory also contains a sample.jsonl with 100 stride-sampled records for quick inspection without loading anything, and manifest.json at the repository root records per-source input/output counts, drops and flag histograms for the build.

Licensing and attribution

The compilation is published under CC-BY-4.0. Individual configs remain governed by their source licenses listed above; the mapa text originates from EUR-LEX and is reused under the European Commission's reuse policy (Decision 2011/833/EU). Please cite the original datasets when you use the corresponding configs: ai4privacy, NVIDIA, Gretel, the privy generator authors, and the MAPA project.