575-lab/kiji-inspector-reviewed-pairs

Kiji PII Detection Training Data

Synthetic multilingual dataset for training PII (Personally Identifiable Information) detection models with token-level entity annotations and coreference resolution.

Dataset Summary

Samples	99,990 (train: 89,991, test: 9,999)
Languages	6 (Dutch, Spanish, German, English, Danish, French)
Countries	20
PII entity types	26
Total entity annotations	814,306 (avg 8.1 per sample)
Coreference clusters	142,142 (99% of samples)
Text length	130–1,149 chars (avg 446)

Usage

from datasets import load_dataset

ds = load_dataset("575-lab/kiji-inspector-reviewed-pairs")

# Access a sample
sample = ds["train"][0]
print(sample["text"])
print(sample["privacy_mask"])   # PII entity annotations
print(sample["coreferences"])   # Coreference clusters
print(sample["language"])       # e.g. "English"
print(sample["country"])        # e.g. "United States"

Schema

Each sample contains:

Column	Type	Description
text	string	Natural language text with embedded PII
privacy_mask	list[{"value": str, "label": str}]	PII entities with their text span and label
coreferences	list[{"mentions": list[str], "entity_type": str, "cluster_id": int}]	Coreference clusters linking mentions of the same entity
language	string	Language of the text
country	string	Country context for the PII (affects address/ID formats)

Example sample

{
  "text": "Contact Dr. Maria Santos at maria.santos@hospital.org or call +1-555-123-4567.",
  "privacy_mask": [
    {"value": "Maria", "label": "FIRSTNAME"},
    {"value": "Santos", "label": "SURNAME"},
    {"value": "maria.santos@hospital.org", "label": "EMAIL"},
    {"value": "+1-555-123-4567", "label": "PHONENUMBER"}
  ],
  "coreferences": [
    {
      "mentions": ["Dr. Maria Santos", "maria.santos"],
      "entity_type": "FIRSTNAME",
      "cluster_id": 0
    }
  ],
  "language": "English",
  "country": "United States"
}

PII Labels

Label	Count
FIRSTNAME	86,136
CITY	80,194
BUILDINGNUM	76,268
SURNAME	74,278
STREET	71,946
ZIP	55,054
STATE	43,194
COUNTRY	22,728
COMPANYNAME	16,036
DATEOFBIRTH	15,114
PHONENUMBER	13,740
EMAIL	13,322
DRIVERLICENSENUM	13,078
SSN	12,886
SECURITYTOKEN	12,816
PASSPORTID	12,702
IBAN	12,688
PASSWORD	12,586
NATIONALID	12,510
TAXNUM	12,346
LICENSEPLATENUM	12,320
IDCARDNUM	12,140
URL	12,096
AGE	11,610
CREDITCARDNUMBER	7,130
USERNAME	6,468

Language Distribution

Language	Samples	%
Spanish	16,846	16.8%
German	16,708	16.7%
Dutch	16,704	16.7%
Danish	16,692	16.7%
English	16,546	16.5%
French	16,494	16.5%

Country Distribution

Country	Samples	%
Denmark	16,692	16.7%
Belgium	11,654	11.7%
Switzerland	8,938	8.9%
Netherlands	8,418	8.4%
Canada	5,924	5.9%
Germany	5,574	5.6%
Austria	5,512	5.5%
Luxembourg	3,304	3.3%
France	3,296	3.3%
Peru	2,890	2.9%
Mexico	2,886	2.9%
Colombia	2,824	2.8%
New Zealand	2,810	2.8%
United Kingdom	2,804	2.8%
Chile	2,790	2.8%
(5 more)	13,674	13.7%

Data Generation

Samples are synthetically generated using LLMs with structured outputs. The generation pipeline:

1. NER generation — LLM produces text with embedded PII and entity annotations
2. Coreference generation — second pass links pronouns and references to their antecedent entities
3. Review (optional) — additional LLM pass validates and corrects annotations
4. Format conversion — samples are converted to a clean, standardized schema

Intended Use

This dataset is designed for training token-classification models that detect and classify PII in text. The coreference annotations enable training models that can also resolve entity mentions (e.g., linking "he" back to "John Smith").

Limitations

• All data is synthetically generated — entity distributions may not match real-world text
• Coreference annotations are LLM-generated and may contain errors
• Address and ID formats are country-specific but may not cover all regional variations