Spyridon/camembert-base-pii-fr

CamemBERT French PII NER

A fine-tuned CamemBERT model for French PII (personally identifiable information) detection, trained to identify 56 entity types for document anonymization / redaction.

Model description

• Base model: almanach/camembert-base (RoBERTa-style, French, 111M params)
• Task: token classification (BIO NER)
• Labels: 56 entity types → 113 BIO labels
• Language: French

Label taxonomy (56 entity types)

Group	Entity types
Names	FIRSTNAME, LASTNAME, MIDDLENAME, PREFIX, USERNAME
Contact	EMAIL, PHONENUMBER, PHONEIMEI, URL, IP, IPV4, IPV6, MAC, USERAGENT
Address	STREET, CITY, STATE, ZIPCODE, BUILDINGNUMBER, COUNTY, SECONDARYADDRESS, NEARBYGPSCOORDINATE, ORDINALDIRECTION
IDs & numbers	ACCOUNTNUMBER, ACCOUNTNAME, IBAN, SSN, CREDITCARDNUMBER, CREDITCARDCVV, CREDITCARDISSUER, VEHICLEVIN, VEHICLEVRM, BIC, MASKEDNUMBER, PIN
Dates	DATE, DOB, TIME
Crypto	BITCOINADDRESS, ETHEREUMADDRESS, LITECOINADDRESS
Personal attributes	AGE, GENDER, SEX, EYECOLOR, HEIGHT, JOBTITLE, JOBAREA, JOBTYPE, COMPANYNAME
Other	PASSWORD, AMOUNT, CURRENCY, CURRENCYCODE, CURRENCYNAME, CURRENCYSYMBOL

Training

• Dataset: French subset of ai4privacy/pii-masking-200k (61,958 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 (same distribution, synthetic): F1 0.94, precision 0.94, recall 0.95

Out-of-distribution — TheoDB/french-pii-eval (8-class taxonomy, span-level exact match):

Metric	Value
Precision	0.51
Recall	0.87
F1	0.65

Per-class F1: private_email 0.93, private_date 0.82, private_phone 0.76, private_address 0.72, account_number 0.68, private_person 0.63, private_url 0.37, secret 0.33.

Note: the model favors recall over precision (over-detects PII), which is the safer direction for anonymization — better to over-redact than leak PII.

Usage

from transformers import pipeline

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

For PDF redaction, see the companion pipeline that maps detected spans back to PDF word boxes and applies redaction annotations.