CamemBERT French PII NER — v2 (8-class)

Second iteration of the French PII detection model. Improvement over v1: a coarse 8-class taxonomy and training on realistic mixed data, which lifted out-of-distribution F1 from 0.65 → 0.74.

Label taxonomy (8 classes)

private_person, private_address, private_email, private_phone, private_url, private_date, account_number, secret.

These match the TheoDB/french-pii-eval / openai/privacy-filter taxonomy. The 56 fine-grained ai4privacy types are collapsed into these 8 (e.g. all of FIRSTNAME/LASTNAME/MIDDLENAME/PREFIX → private_person).

Training

• Base model: almanach/camembert-base
• Data: TheoDB/french-pii-eval train + val (57.7K) mixed with the French subset of ai4privacy/pii-masking-200k (61.9K), mapped to 8 classes and deduplicated (~67K unique examples)
• Hyperparameters: 3 epochs, batch size 16 (gradient accumulation 2), learning rate 5e-5, max sequence length 128, weight decay 0.01
• Hardware: single NVIDIA A10G

Evaluation

Held-out split (in-distribution): F1 0.985.

Out-of-distribution, TheoDB/french-pii-eval test (2,500 examples), span-level exact match:

Per-class F1 (v2): private_email 0.99, private_address 0.88, account_number 0.84, private_date 0.84, private_phone 0.80, private_person 0.62, secret 0.44, private_url 0.41.

Threshold: F1 is flat (0.737–0.739) across thresholds 0.3–0.9. For anonymization, favor a lower threshold (higher recall, e.g. 0.3–0.5); for less over-redaction, use 0.7–0.8.

Usage

from transformers import pipeline

nlp = pipeline("token-classification", model="Spyridon/camembert-pii-fr-v2",
               aggregation_strategy="simple")
nlp("Contactez Jean Dupont au 06 12 34 56 78 ou jean.dupont@email.fr")