PII-SCOPE

PII-SCOPE is a benchmark for evaluating models that detect personally identifiable information (PII) in text. Its documents are long, realistic and densely packed with PII of many types, so it is harder than most public PII benchmarks. It is released with the Protegrity-PII Technical Report (https://arxiv.org/abs/2610.30518).

Key facts

• 8,030 documents with 114,887 labeled PII spans across 69 entity types.
• Rich documents: on average 14.31 PII spans and 8.17 different entity types per document, the highest of the benchmarks compared in the paper. Mean document length is about 1,800 characters.
• Many scenarios: domains include retail, healthcare, government, technology, finance, insurance, human resources, legal, telecom, education, hospitality, travel, energy, manufacturing, real estate and logistics.
• Many formats: emails, letters, medical records, resumes, receipts, invoices, chat logs, support tickets, social media and forum posts, forms and meeting notes. About 95% of documents are unstructured text and 5% are structured.
• Language: English.
• Fully synthetic: generated with Protegrity's proprietary LLM-based data generation pipeline. All PII values are fake; the dataset contains no real personal data.

Entity types

ACCESS_SECURITY_CODE, ACCOUNT_BALANCE, ACCOUNTNAME, ACCOUNTNUMBER, AGE, AMOUNT,
BIC, BUILDING, CARD_EXPIRATION_DATE, CARD_SERVICE_CODE, CARD_TRACK_DATA, CCN,
CITY, COMPANYNAME, COUNTRY, COUNTY, CREDIT_SCORE, CREDITCARDCVV,
CRIMINAL_RECORD, CRYPTO_ADDRESS, CURRENCY, DATE, DEVICE_IDENTIFIER, DRIVERLICENSE,
EMAIL, FAX_NUMBER, FINGERPRINT_DATA, GENDER, GENETIC_DATA, GEOCOORD,
HEALTH_CONDITION, HEALTH_PLAN_BENEFICIARY_NUMBER, IBAN, IDCARD, IMMIGRATION_STATUS, INCOME,
IP, LOGIN_CREDENTIAL, MAC, MEDICAL_RECORD_NUMBER, MOTHERS_MAIDEN_NAME, NAME,
NEURAL_DATA, PASSPORT, PASSWORD, PHONE, PIN, PLACE_OF_BIRTH,
POLITICAL_OPINION, PROFESSIONAL_LICENSE_NUMBER, RACE_ETHNICITY, RELIGIOUS_BELIEF, RETINA_SCAN_DATA, SECADDRESS,
SEX_LIFE_DATA, SEXUAL_ORIENTATION, SSN, STATE, STREET, TIME,
TITLE, TRACKING_IDENTIFIER, TRADE_UNION_MEMBERSHIP, TRANSACTION_HISTORY, URL, USERNAME,
VEHICLE_IDENTIFICATION_NUMBER, VEHICLE_REGISTRATION, ZIPCODE

Row format

The dataset is one CSV file, pii_scope.csv, with one document per row and three columns:

• source_text: the document text.
• privacy_mask: a JSON list of PII spans. Each span has value (the PII text), label (one of the 69 entity types), and start and end (character offsets into source_text; end is exclusive, so source_text[start:end] == value).
• language: the document language.

Example row:

{
  "source_text": "Just decluttering my place and decided to give away some furniture for free! 🛋️💨\nIf you need a cozy couch, a solid dining table, or even a cute bookshelf—hit me up! Pickup is easy: come by 776 Martinez Road Suite 369 on 10/23/2003 between 10 AM and 4 PM. Reach out to me at +44 4042 670714 or drop me a line at john_sandoval@phillipsllc.fr. First-come, first-served—don’t wait too long, these won’t last! 😄\nThanks, Justin Obrien, for helping keep good stuff in good hands!",
  "privacy_mask": [
    {"value": "776 Martinez Road Suite 369", "label": "STREET", "start": 189, "end": 216},
    {"value": "10/23/2003", "label": "DATE", "start": 220, "end": 230},
    {"value": "10 AM", "label": "TIME", "start": 239, "end": 244},
    {"value": "4 PM", "label": "TIME", "start": 249, "end": 253},
    {"value": "+44 4042 670714", "label": "PHONE", "start": 274, "end": 289},
    {"value": "john_sandoval@phillipsllc.fr", "label": "EMAIL", "start": 311, "end": 339},
    {"value": "Justin Obrien", "label": "NAME", "start": 415, "end": 428}
  ],
  "language": "English"
}

privacy_mask is stored as a JSON string. To read it:

import json
import pandas as pd

df = pd.read_csv("pii_scope.csv")
spans = json.loads(df.loc[0, "privacy_mask"])
print([(df.loc[0, "source_text"][s["start"]:s["end"]], s["label"]) for s in spans])

Use cases

• Benchmarking PII detectors: compare models on long, dense documents with a fine-grained taxonomy. The paper scores exact span matches, with and without the entity type (typed and untyped precision, recall and F1).
• Testing taxonomy coverage: measure how much PII a detector misses because its taxonomy does not support an entity type.
• Evaluating fine-tuned models: check that a model fine-tuned on new data still detects the full range of entity types.
• Redaction and anonymization research: study how detection errors affect downstream redaction of documents such as emails, medical records and financial forms.

Licensing

PII-SCOPE is released under LICENSE.md for research and non-commercial use only. Commercial use is not permitted. For questions, contact hardeep.bassi@protegrity.com or ave.gatton@protegrity.com.

Citation

If you use PII-SCOPE, please cite the Protegrity-PII Technical Report (https://arxiv.org/abs/2610.30518).