PII Detection Dataset

This dataset is derived from gretelai/synthetic_pii_finance_multilingual, filtered to English samples with PII labels consolidated into 11 standardized categories for evaluating PII detection metrics.

Dataset Creation

import ast
from datasets import load_dataset

# Load source dataset and filter to English
ds = load_dataset("gretelai/synthetic_pii_finance_multilingual", split="test")
ds = ds.filter(lambda x: x["language"] == "English")

# Map fine-grained labels to consolidated PII classes
LABEL_MAPPING = {
    'bban': 'account_info',
    'date_time': None,
    'email': 'email',
    'password': 'password',
    'ssn': 'ssn',
    'api_key': 'password',
    'time': None,
    'local_latlng': None,
    'user_name': 'username',
    'credit_card_security_code': 'credit_card_info',
    'ipv4': 'network_info',
    'passport_number': None,
    'swift_bic_code': 'account_info',
    'first_name': 'name',
    'date': None,
    'ipv6': 'network_info',
    'iban': 'account_info',
    'customer_id': None,
    'company': None,
    'account_pin': 'password',
    'name': 'name',
    'employee_id': None,
    'credit_card_number': 'credit_card_info',
    'bank_routing_number': 'account_info',
    'street_address': 'address',
    'driver_license_number': None,
    'last_name': 'name',
    'date_of_birth': 'date_of_birth',
    'phone_number': 'phone_number',
}

def has_non_none_labels(example):
    spans = ast.literal_eval(example['pii_spans'])
    labels = [item['label'] for item in spans]
    if not labels:
        return True
    return any(LABEL_MAPPING.get(label) is not None for label in labels)

def transform(example):
    spans = ast.literal_eval(example['pii_spans'])
    labels = [item['label'] for item in spans]
    converted = list(set(LABEL_MAPPING[l] for l in labels if LABEL_MAPPING.get(l)))
    return {'text': example['generated_text'], 'labels': converted}

ds = ds.filter(has_non_none_labels)
ds = ds.map(transform, remove_columns=ds.column_names)

Labels not in this mapping (e.g., date_time, passport_number, company) are excluded. Samples containing only excluded labels are filtered out.