mmbert-pii-detector-merged

mmBERT PII Detector (Merged for Rust)

This is a merged mmBERT model for PII (Personally Identifiable Information) detection, optimized for Rust inference using the candle framework.

Model Details

• Base Model: jhu-clsp/mmBERT-base
• Task: Token classification for PII detection
• Languages: Multilingual (1800+ languages)
• Training: LoRA fine-tuned then merged with base model
• Inference: Optimized for Rust candle-binding

PII Types Detected

• Names (PERSON)
• Email addresses (EMAIL)
• Phone numbers (PHONE)
• Addresses (ADDRESS)
• Credit card numbers (CREDIT_CARD)
• Social Security Numbers (SSN)
• Dates (DATE)
• And more...

Usage with Rust (candle-binding)

use candle_semantic_router::model_architectures::traditional::TraditionalModernBertTokenClassifier;

let classifier = TraditionalModernBertTokenClassifier::new("path/to/model", true)?;
let results = classifier.classify_tokens("Contact [email protected] or call 555-1234")?;
for (token, label, confidence) in results {
    if label != "O" {
        println!("PII: {} -> {} ({:.1}%)", token, label, confidence * 100.0);
    }
}

Training Configuration

• LoRA Rank: 32
• LoRA Alpha: 64
• Epochs: 10
• Batch Size: 64
• Learning Rate: 2e-5

License

Apache 2.0

---

Maintenance

• Repo README updated by automation (timestamp: 2026-01-30 UTC).