mmbert-pii-detector-lora

mmBERT PII Detector (LoRA Adapter)

A multilingual token classification model for Personal Identifiable Information (PII) detection using mmBERT with LoRA adapters.

Model Description

This model detects and classifies PII entities at the token level using BIO tagging:

• PERSON: Names of individuals
• EMAIL_ADDRESS: Email addresses
• PHONE_NUMBER: Phone numbers
• STREET_ADDRESS: Physical addresses
• CREDIT_CARD: Credit card numbers
• US_SSN: US Social Security Numbers
• And more...

Performance

Training Details

• Base Model: jhu-clsp/mmBERT-base
• LoRA Rank: 32
• LoRA Alpha: 64
• Task: Token Classification (BIO tagging)
• Dataset: Microsoft Presidio Research Dataset

Usage

from peft import PeftModel
from transformers import AutoModelForTokenClassification, AutoTokenizer

# Load model
base_model = AutoModelForTokenClassification.from_pretrained(
    "jhu-clsp/mmBERT-base", num_labels=35  # O + 17 entity types × 2 (B/I)
)
model = PeftModel.from_pretrained(base_model, "vllm-sr/mmbert-pii-detector-lora")
tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/mmBERT-base")

# Detect PII
text = "Contact John Smith at [email protected] or call 555-123-4567"
inputs = tokenizer(text.split(), is_split_into_words=True, return_tensors="pt")
outputs = model(**inputs)
predictions = outputs.logits.argmax(-1)

# Map predictions to labels
# ... (see full example in repository)

Supported PII Types

Use Cases

• Privacy Compliance: GDPR, CCPA compliance
• Data Masking: Redact PII before processing
• LLM Safety: Prevent PII leakage in LLM outputs
• Document Processing: Automated PII detection in documents

Part of vLLM Semantic Router

This model is part of the vLLM Semantic Router project.

License

Apache 2.0