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
