jamon8888/french-pii-legal-ner-base

---

language: fr tags: - gliner - ner - pii - gdpr - legal - french license: apache-2.0

🇫🇷 French Legal PII NER — GLiNER2 Base (Lightweight)

Lightweight fine-tuned GLiNER2 Base for French PII detection in legal documents. Optimized for laptop/edge deployment. GDPR & EU AI Act compliant.

Evaluation Results

Model	F1	Precision	Recall
This model (GLiNER2 Base, fine-tuned)	0.530	0.612	0.528
GLiNER2 Base (no fine-tuning)	0.495	0.577	0.472
GPT-4o	0.836	0.863	0.844

Why use this over the Large version?

• ✅ 4x smaller (~680 MB vs ~2.2 GB)
• ✅ 3x faster on CPU
• ✅ Only 2.5% F1 gap vs Large
• ✅ Runs on any modern laptop with 8GB RAM

Supported Entity Types (33)

nom_personne, prenom, date_naissance, lieu_naissance, adresse, numero_telephone, email, numero_securite_sociale, numero_passeport, numero_carte_identite, numero_siret, organisation, tribunal, numero_affaire, avocat, notaire, juge, date, lieu, nationalite, profession, salaire, numero_compte_bancaire, plaque_immatriculation, donnee_sante, donnee_biometrique, donnee_genetique, opinion_politique, religion, origine_ethnique, orientation_sexuelle, condamnation_penale, base_legale_traitement

Usage

```bash pip install gliner ```

```python from gliner import GLiNER

model = GLiNER.from_pretrained( "jamon8888/french-pii-legal-ner-base", base_model="fastino/gliner2-base-v1" )

text = "M. Jean Dupont, né le 12/03/1985 à Paris, domicilié au 15 rue de la Paix..." labels = ["nom_personne", "date_naissance", "adresse", "donnee_sante"]

entities = model.predict_entities(text, labels, threshold=0.5) for entity in entities: print(f"{entity['label']}: {entity['text']}") ```

Training Details

• Base model: fastino/gliner2-base-v1
• Training examples: 995
• Epochs: 8
• Learning rate: 5e-5
• Document types: contrats, assignations, jugements, actes notariés, rapports médico-légaux, dossiers RH
• Trained with: Pioneer by Fastino Labs