# Tegmen

A high-performance on-premise PII detection and masking solution

## Overview

Tegmen is a production-ready token classification system designed for identifying and masking personally identifiable information (PII) in text data. Built for high-throughput data sanitization workflows, it offers on-premise deployment capabilities with enterprise-grade performance.

## Key Features

- **On-Premise Deployment**: Run entirely within your infrastructure
- **Lightweight Architecture**: Optimized for edge deployment
- **Fine-Tunable**: Easily adapt to your specific data distributions
- **Long Context Support**: Process documents up to 128,000 tokens
- **Configurable Detection**: Tune precision/recall tradeoffs

## Supported PII Categories

The model detects 8 categories of sensitive information:

| Category | Description |
|----------|-------------|
| `account_number` | Financial account identifiers |
| `private_address` | Physical and mailing addresses |
| `private_email` | Email addresses |
| `private_person` | Personal names |
| `private_phone` | Phone numbers |
| `private_url` | URLs and web addresses |
| `private_date` | Birth dates and personal dates |
| `secret` | API keys, passwords, credentials |

## Installation

```bash
pip install transformers torch
```

## Quick Start

### Using the Pipeline API

```python
from transformers import pipeline

detector = pipeline("token-classification", model="comethrusws/tegmen", aggregation_strategy="simple")

text = "Contact John Smith at john.smith@email.com"
results = detector(text)

for item in results:
    print(f"Found: {item['word']} ({item['entity_group']})")
```

### Using the Model Directly

```python
import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("comethrusws/tegmen")
model = AutoModelForTokenClassification.from_pretrained("comethrusws/tegmen")

text = "My name is Alice and my email is alice@example.com"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)

predictions = outputs.logits.argmax(dim=-1)
labels = [model.config.id2label[p.item()] for p in predictions[0]]
print(labels)
```

## Performance Specifications

- **Architecture**: Transformer encoder
- **Parameters**: 1.5B total / 50M active
- **Context Window**: 128,000 tokens
- **Output Format**: BIOES span tagging

## License

Apache License 2.0

## Support

For enterprise support, contact <a href="https://sagea.space">SAGEA</a>.