sentinel-pii-sdk 0.1.0

Python SDK for Sentinel PII Redaction - State-of-the-art PII detection and redaction using fine-tuned Granite models

Sentinel PII SDK

State-of-the-art PII detection and redaction using the Sentinel model

Sentinel PII SDK is a Python library for identifying and redacting Personally Identifiable Information (PII) in text.

Features

• High-accuracy PII detection (95%+ recall)
• Multiple handling modes: TAG, REDACT, or REPLACE
• Batch processing support

Installation

From PyPI

pip install sentinel-pii-sdk

With faker support for REPLACE mode:

pip install 'sentinel-pii-sdk[faker]'

From Source

git clone https://github.com/cernis-intelligence/sentinel-pii-sdk.git
cd sentinel-pii-sdk
pip install -e .

Quick Start

from sentinel_pii import SentinelPIIRedactor

# Initialize (model loads from HuggingFace on first use)
redactor = SentinelPIIRedactor()

# Detect PII in text
text = "My name is John Smith and my email is [e-mail]"
result = redactor.redact_text(text)
print(result)
# Output: "My name is [PERSON_NAME] and my email is [EMAIL_ADDRESS]"

Usage Examples

Basic PII Detection

from sentinel_pii import SentinelPIIRedactor, PIIHandlingMode

redactor = SentinelPIIRedactor()

text = "Contact John Smith at [e-mail] or call [phone]"

# TAG mode - Show PII categories
result = redactor.redact_text(text, mode=PIIHandlingMode.TAG)
print(result)
# "Contact [PERSON_NAME] at [EMAIL_ADDRESS] or call [PHONE_NUMBER]"

# REDACT mode - Same as TAG
result = redactor.redact_text(text, mode=PIIHandlingMode.REDACT)
print(result)
# "Contact [PERSON_NAME] at [EMAIL_ADDRESS] or call [PHONE_NUMBER]"

# REPLACE mode - Replace with fake data (requires faker)
result = redactor.redact_text(text, mode=PIIHandlingMode.REPLACE)
print(result)
# "Contact Jane Doe at [e-mail] or call [phone]"

Batch Processing

from sentinel_pii import detect_pii_batch, PIIHandlingMode

documents = [
    "My email is [e-mail]",
    "Patient DOB: 1990-05-15, diagnosed with diabetes"
]

results = detect_pii_batch(documents, mode=PIIHandlingMode.TAG)
for result in results:
    print(result)

Dataset Cleaning

from sentinel_pii import clean_dataset, PIIHandlingMode

# Clean a JSONL dataset file
clean_dataset(
    input_filename="input_data.jsonl",
    output_filename="output_data.jsonl",
    mode=PIIHandlingMode.TAG
)

Supported PII Categories

The Sentinel model detects 20+ PII categories:

Identity: PERSON_NAME, USERNAME, AGE, GENDER, DEMOGRAPHIC_GROUP

Contact: EMAIL_ADDRESS, PHONE_NUMBER, STREET_ADDRESS, CITY, STATE, POSTCODE, COUNTRY

Dates: DATE, DATE_OF_BIRTH

ID Numbers: PERSONAL_ID, PASSPORT, DRIVERLICENSE

Financial: CREDIT_CARD_INFO, BANKING_NUMBER

Security: PASSWORD, SECURE_CREDENTIAL

Medical: MEDICAL_CONDITION

Other: ORGANIZATION_NAME, DOMAIN_NAME, NATIONALITY, RELIGIOUS_AFFILIATION

API Reference

SentinelPIIRedactor

Main class for PII detection.

redactor = SentinelPIIRedactor(pii_categories=None)

Parameters:

• pii_categories (optional): Custom PII categories string

Methods:

• redact_text(text, mode=PIIHandlingMode.TAG, locale="en_US") - Process single text
• detect_pii(documents, mode=PIIHandlingMode.TAG, locale="en_US", show_progress=True) - Process list of documents

Utility Functions

• detect_pii_batch(documents, mode=PIIHandlingMode.TAG, locale="en_US") - Batch processing
• clean_dataset(input_filename, output_filename, mode=PIIHandlingMode.TAG, locale="en_US") - Clean JSONL files

PIIHandlingMode

Enum for handling modes:

• PIIHandlingMode.TAG - Show PII categories in brackets
• PIIHandlingMode.REDACT - Same as TAG
• PIIHandlingMode.REPLACE - Replace with fake data (requires faker)

Model Information

• Model: cernis-intelligence/sentinel on HuggingFace
• Performance: 95%+ recall, ~100 docs/min on GPU
• License: Apache 2.0

Requirements

• Python >= 3.9
• transformers >= 4.36.0
• torch >= 2.0.0
• accelerate >= 0.20.0
• tqdm >= 4.65.0
• faker >= 20.0.0 (optional, for REPLACE mode)

Examples

The examples/ directory contains working sample scripts:

# Basic single-text PII detection
python3.11 examples/basic_usage.py

# Process multiple documents at once
python3.11 examples/batch_processing.py

# Clean JSONL dataset files
python3.11 examples/dataset_cleaning.py

# Validate package structure (no model download)
python3.11 examples/test_all_examples.py

You can also use the included sample_data.jsonl for testing:

from sentinel_pii import clean_dataset, PIIHandlingMode

clean_dataset(
    "examples/sample_data.jsonl",
    "output.jsonl",
    mode=PIIHandlingMode.TAG
)

Contributing

Contributions welcome! Please submit a Pull Request.

License

Apache 2.0 License - see LICENSE file for details.

Support

• HuggingFace: cernis-intelligence/sentinel
• Issues: GitHub Issues

Acknowledgments

• Built on IBM Granite 4.0
• Training data from AI4Privacy