DataFog PII-Small-EN (v1.4)

A compact PII (Personally Identifiable Information) Named Entity Recognition model for English text.

Architecture

• Backbone: DeBERTa-v3-xsmall (22M params)
• Head: CharCNN + CRF
• Total params: ~45M
• Entity types: 41 PII categories across 4 sensitivity tiers

Performance (v1.4, Epoch 3)

Tier	Recall	Target
T1 (Critical: SSN, Credit Card, etc.)	0.814	>=0.98
T2 (High: Person, Email, Phone, etc.)	0.937	>=0.95
T3 (Moderate: Username, Date, Location, etc.)	0.945	>=0.90
T4 (Domain-specific: Employee ID, Crypto, etc.)	0.937	>=0.85
Overall F1	0.889

Training Data

• AI4Privacy (~43K examples, English subset)
• NVIDIA Nemotron-PII (100K examples)
• Gretel Synthetic PII Finance (26K examples)
• Gretel PII Masking EN v1 (50K examples)
• Synthetic data for rare entity types (22K examples)
• Total: ~241K examples

Entity Types (41)

Tier 1 — Critical PII

SSN, CREDIT_CARD, BANK_ACCOUNT, PASSPORT_NUMBER, DRIVERS_LICENSE, TAX_ID

Tier 2 — High Sensitivity

PERSON, EMAIL, PHONE, DATE_OF_BIRTH, STREET_ADDRESS, IP_ADDRESS

Tier 3 — Moderate Sensitivity

USERNAME, DATE, LOCATION, ORGANIZATION, URL, LICENSE_PLATE, AGE, NATIONALITY, GENDER, RELIGION, MARITAL_STATUS

Tier 4 — Domain-Specific

MEDICAL_RECORD, EMPLOYEE_ID, STUDENT_ID, ACCOUNT_NUMBER, PIN, PASSWORD, BIOMETRIC, VEHICLE_ID, DEVICE_ID, CRYPTO_WALLET, IBAN, SWIFT_CODE, INSURANCE_NUMBER, SALARY, CRIMINAL_RECORD, POLITICAL_AFFILIATION, SEXUAL_ORIENTATION, HEALTH_CONDITION

License

Apache 2.0