OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1

Portuguese PII Detection Model | 568M Parameters | Open Source

Model Description

OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1 is a transformer-based token classification model fine-tuned for Personally Identifiable Information (PII) detection in Portuguese text. This model identifies and classifies 54 types of sensitive information including names, addresses, social security numbers, medical record numbers, and more.

Key Features

• Portuguese-Optimized: Specifically trained on Portuguese text for optimal performance
• High Accuracy: Achieves strong F1 scores across diverse PII categories
• Comprehensive Coverage: Detects 55+ entity types spanning personal, financial, medical, and contact information
• Privacy-Focused: Designed for de-identification and compliance with GDPR and other privacy regulations
• Production-Ready: Optimized for real-world text processing pipelines

Performance

Evaluated on the Portuguese test split (AI4Privacy + synthetic data):

Top 10 Portuguese PII Models

Supported Entity Types

This model detects 54 PII entity types organized into categories:

Identifiers (22 types) Personal Info (11 types) Contact Info (2 types) Location (9 types) Organization (3 types) Financial (5 types) Temporal (2 types)

Usage

Quick Start

from transformers import pipeline

# Load the PII detection pipeline
ner = pipeline("ner", model="OpenMed/OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1", aggregation_strategy="simple")

text = """
Paciente Carlos Mendes (nascido em 15/03/1985, CPF: 987.654.321-00) foi atendido hoje.
Contato: carlos.mendes@email.pt, Telefone: +351 912 345 678.
Endereço: Avenida da Liberdade 42, 1250-096 Lisboa.
"""

entities = ner(text)
for entity in entities:
    print(f"{entity['entity_group']}: {entity['word']} (score: {entity['score']:.3f})")

De-identification Example

def redact_pii(text, entities, placeholder='[REDACTED]'):
    """Replace detected PII with placeholders."""
    # Sort entities by start position (descending) to preserve offsets
    sorted_entities = sorted(entities, key=lambda x: x['start'], reverse=True)
    redacted = text
    for ent in sorted_entities:
        redacted = redacted[:ent['start']] + f"[{ent['entity_group']}]" + redacted[ent['end']:]
    return redacted

# Apply de-identification
redacted_text = redact_pii(text, entities)
print(redacted_text)

Batch Processing

from transformers import AutoModelForTokenClassification, AutoTokenizer
import torch

model_name = "OpenMed/OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1"
model = AutoModelForTokenClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

texts = [
    "Paciente Carlos Mendes (nascido em 15/03/1985, CPF: 987.654.321-00) foi atendido hoje.",
    "Contato: carlos.mendes@email.pt, Telefone: +351 912 345 678.",
]

inputs = tokenizer(texts, return_tensors='pt', padding=True, truncation=True)
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.argmax(outputs.logits, dim=-1)

Training Details

Dataset

This model was trained on a combination of:

• AI4Privacy PII Masking 200K: Multilingual base dataset (200K records across 8 languages)

• NVIDIA Nemotron-PII: Seed dataset for synthetic data generation

• Synthetic Portuguese Data: ~80K high-quality samples generated with locale-specific formatting (CPF format, +55 phones, Brazilian names, R$ currency)

• Format: BIO-tagged token classification

• Labels: 76 BIO tags (54 entity types)

Training Configuration

• Max Sequence Length: 512 tokens
• Epochs: 3
• Framework: Hugging Face Transformers + Trainer API

Intended Use & Limitations

Intended Use

• De-identification: Automated redaction of PII in Portuguese clinical notes, medical records, and documents
• Compliance: Supporting GDPR, and other privacy regulation compliance
• Data Preprocessing: Preparing datasets for research by removing sensitive information
• Audit Support: Identifying PII in document collections

Limitations

Important: This model is intended as an assistive tool, not a replacement for human review.

• False Negatives: Some PII may not be detected; always verify critical applications
• Context Sensitivity: Performance may vary with domain-specific terminology
• Language: Optimized for Portuguese text; may not perform well on other languages

Citation

@misc{openmed-pii-2026,
  title = {OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1: Portuguese PII Detection Model},
  author = {OpenMed Science},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/OpenMed/OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1}
}

Links

• Organization: OpenMed