SoelMgd/bert-pii-detection

BERT PII Detection Model

Fine-tuned DistilBERT model for Personal Identifiable Information (PII) detection and classification.

Model Details

• Base Model: distilbert-base-uncased
• Task: Token Classification (Named Entity Recognition)
• Languages: English
• License: MIT
• Fine-tuned on: AI4Privacy PII-42k dataset

Supported PII Entity Types

This model can detect 56 different types of PII entities including:

Personal Information:

• FIRSTNAME, LASTNAME, MIDDLENAME
• EMAIL, PHONENUMBER, USERNAME
• DATE, TIME, DOB, AGE

Address Information:

• STREET, CITY, STATE, COUNTY
• ZIPCODE, BUILDINGNUMBER
• SECONDARYADDRESS

Financial Information:

• CREDITCARDNUMBER, CREDITCARDISSUER, CREDITCARDCVV
• ACCOUNTNAME, ACCOUNTNUMBER, IBAN, BIC
• AMOUNT, CURRENCY, CURRENCYCODE, CURRENCYSYMBOL

Identification:

• SSN, PIN, PASSWORD
• IP, IPV4, IPV6, MAC
• ETHEREUMADDRESS, BITCOINADDRESS, LITECOINADDRESS

Professional Information:

• JOBTITLE, JOBTYPE, JOBAREA, COMPANYNAME

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline

# Load model and tokenizer
model_name = "SoelMgd/bert-pii-detection"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

# Create NER pipeline
ner_pipeline = pipeline(
    "ner", 
    model=model, 
    tokenizer=tokenizer,
    aggregation_strategy="simple"
)

# Example usage
text = "Hi, my name is John Smith and my email is john.smith@company.com"
entities = ner_pipeline(text)
print(entities)

Training Data

• Dataset: AI4Privacy PII-200k
• Size: ~209k examples
• Languages: English, French, German, Italian (this model: English only)
• Entity Types: 56 different PII categories

Performance

The model achieves high performance on PII detection tasks with good precision and recall across different entity types.

Intended Use

This model is designed for:

• PII detection and masking in text
• Privacy compliance applications
• Data anonymization pipelines
• Content moderation systems

Limitations

• Trained primarily on English text
• May not generalize to domain-specific jargon
• Performance may vary on very short or very long texts
• Should be validated on your specific use case