OskarBartoszyk/PLAnonimizer

HerBERT NER – RODO / GDPR Anonymization (PL)

Fine-tuned HerBERT model for Named Entity Recognition (NER) focused on automatic anonymization of RODO / GDPR-sensitive data in Polish-language texts, with emphasis on medical and administrative documents.

The model is designed to detect personal data, sensitive attributes, identifiers, and contact information to support downstream anonymization pipelines.

---

✨ Use case

• Anonymization of medical documentation
• GDPR / RODO compliance
• Text preprocessing before analytics or ML
• Research on privacy-preserving NLP for Polish

---

🧠 Model details

• Base model: HerBERT (Polish BERT)
• Task: Token classification (NER)
• Labeling scheme: BIO
• Training: Supervised fine-tuning
• Framework: 🤗 Transformers + PyTorch

---

🏷️ Supported entity classes

Personal identifiers

• NAME – first names
• SURNAME – last names
• AGE – age
• DATE_OF_BIRTH – date of birth
• DATE – other dates identifying events
• SEX – sex / gender (explicit)
• RELIGION – religion
• POLITICAL_VIEW – political views
• ETHNICITY – ethnicity / nationality
• SEXUAL_ORIENTATION – sexual orientation
• HEALTH – health-related information
• RELATIVE – family relations revealing identity

Contact & location

• CITY – city (general context)
• ADDRESS – full address
• EMAIL – email addresses
• PHONE – phone numbers

Identifiers

• PESEL – Polish national ID number
• DOCUMENT_NUMBER – ID / passport / license numbers

Professional & education

• COMPANY – employer
• SCHOOL_NAME – school name
• JOB_TITLE – job position

Financial

• BANK_ACCOUNT – bank account numbers
• CREDIT_CARD_NUMBER – payment cards

Digital identifiers

• USERNAME – usernames / logins
• SECRET – passwords, API keys

---

📊 Evaluation

• Validation F1-score: ~0.80
• Precision-oriented configuration (privacy-first)
• Evaluation performed on held-out validation set

Note: For anonymization tasks, recall is prioritized to minimize data leakage.

---