This model performs Named Entity Recognition (NER) to detect sensitive personal and financial information in text. It uses BIO tagging (Begin / Inside / Outside) and supports multilingual input, with strong performance on French, English, and Tunisian Arabic-Latin mixed text.

Detected Entities

Label	Description	Examples
`IBAN`	International Bank Account Number	`TN59 1000 6035 1835 9848 3270`, `FR76 3000 6000 0112...`
`CARD`	Credit / Debit card number	`4532 1234 5678 9012`, `5412 7534 2589 6315`
`PHONE`	Phone number (international formats)	`+216 23 456 789`, `0023 4567890`
`EMAIL`	Email address	`support@yalen.tn`, `ahmed.ben@gmail.com`
`DATE`	Date (any format)	`15/03/2025`, `14 janvier 1990`, `12/2028`

Quick Start

python

from transformers import pipeline

ner = pipeline(
    "ner",
    model="yalen-ai/xlm_roberta_yalen",
    aggregation_strategy="simple"
)

text = "Contact: support@yalen.tn | Tel: +216 23 456 789 | IBAN: TN59 1000 6035 1835 9848 3270"
results = ner(text)

for entity in results:
    print(f"[{entity['entity_group']}] '{entity['word']}' (score: {entity['score']:.3f})")

Output:

[EMAIL] 'support@yalen.tn' (score: 1.000)
[PHONE] '+216 23 456 789' (score: 1.000)
[IBAN] 'TN59 1000 6035 1835 9848 3270' (score: 1.000)

Installation

pip install transformers torch

Usage Examples

Detect all PII in a document

python

from transformers import pipeline

ner = pipeline("ner", model="yalen-ai/xlm_roberta_yalen", aggregation_strategy="simple")

examples = [
    "Rendez-vous le 15/03/2025, contactez ahmed.ben@gmail.com",
    "Carte bancaire: 4532 1234 5678 9012, expire 12/2028",
    "Virement IBAN FR76 3000 6000 0112 3456 7890 189",
    "Appelez le +216 98 765 432 ou +33 1 23 45 67 89",
]

for text in examples:
    print(f"\nTexte: {text}")
    for e in ner(text):
        print(f"  [{e['entity_group']}] {e['word'].strip('.,;')} ({e['score']:.3f})")

Post-processing tip

The tokenizer may include trailing punctuation in detected spans. Clean it up:

python

def clean_entity(word: str) -> str:
    return word.strip(".,;:!? ")

results = ner("Email: contact@yalen.tn, Tel: +216 23 000 000")
cleaned = [{**e, "word": clean_entity(e["word"])} for e in results]

With explicit label mapping

python

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("yalen-ai/xlm_roberta_yalen")
    model = AutoModelForTokenClassification.from_pretrained("yalen-ai/xlm_roberta_yalen")

text = "IBAN: TN59 1000 6035 1835 9848 3270"
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)

predictions = torch.argmax(outputs.logits, dim=2)
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
labels = [model.config.id2label[p.item()] for p in predictions[0]]

for token, label in zip(tokens, labels):
    if label != "O":
        print(f"{token:20} → {label}")

Model Architecture

Parameter	Value
Base model	`xlm-roberta-base`
Architecture	`XLMRobertaForTokenClassification`
Hidden size	768
Attention heads	12
Hidden layers	12
Max tokens	512
Vocab size	250,002
Labels	11 (O + 5×BIO)
Model size	~1.06 GB

Label Map

json

{
  "0": "O",
  "1": "B-IBAN",  "2": "I-IBAN",
  "3": "B-CARD",  "4": "I-CARD",
  "5": "B-PHONE", "6": "I-PHONE",
  "7": "B-EMAIL", "8": "I-EMAIL",
  "9": "B-DATE",  "10": "I-DATE"
}

Intended Use

Privacy compliance (GDPR, Tunisian Data Protection Law)
Document redaction — anonymize sensitive documents before sharing
Data loss prevention (DLP) — detect accidental PII leaks in logs or messages
Financial document processing — extract IBAN/card numbers for validation
Healthcare & insurance — detect dates and contact information

Limitations

Maximum input length: 512 tokens (long documents should be split by sentence or paragraph)
Trained primarily on French and English data — performance may vary on other languages
Card number detection works best with standard spacing formats (XXXX XXXX XXXX XXXX)
Trailing punctuation may be included in token spans — apply post-processing (.strip(".,;"))

Training Data

The model was fine-tuned on a dataset composed of:

Source	Type	Volume	Languages
Synthetically generated PII	IBAN, CARD, PHONE, EMAIL, DATE	~50,000 examples	FR, EN, AR-latin
Filtered CommonCrawl corpus	Neutral text contexts	~20,000 examples	Multilingual
Anonymized internal Yalen data	Tunisian/French formats	~5,000 examples	FR, AR-latin

No real patient data was used in training.

All personally identifiable information in training data is either synthetic or fully anonymized prior to use.

Intended Use

Primary Use Case

Validation of PII detections in the Yalen Artemesia Desktop Guardian compliance audit tool. The model acts as a second-pass validator after regex-based detection, providing confidence scores to reduce false positives.

Inference Pipeline

Raw text (file content)
      ↓
Regex patterns (email, IBAN, CARD, PHONE, DATE)  → raw findings
      ↓
XLM-RoBERTa (this model)
  - context window: ±100 chars around each detected value
  - BIO token classification
  - confidence score per entity
      ↓
Threshold: 0.55
  ≥ 0.55 → slm_confirmed=True  → included in audit report
  < 0.55 → slm_confirmed=False → flagged as uncertain (not silently dropped)

Out-of-Scope Uses

Real-time surveillance or profiling of individuals
Automated decisions affecting individuals' rights
Use outside of compliance audit contexts

Bias, Risks, and Limitations

Known Biases

Bias	Risk	Mitigation
Linguistic bias	Lower performance on non-latin Arabic scripts	Documented limitation; AR-latin covered in training data
IBAN format bias	Stronger on FR/TN formats	Synthetic data covers 30+ country formats
Threshold sensitivity	Score 0.40–0.55 edge cases	Uncertain findings preserved in report with flag, not silently dropped
Training data distribution	Overrepresentation of FR/TN data	Stated limitation; multilingual XLM-R base partially compensates

Risk Mitigation

Graceful degradation: if the model is unavailable, the system falls back to regex-only detection with slm_score=1.0 — no silent failures
Human in the loop: the tool produces a consultative audit report; no automated decision affecting individuals is made
Transparency: every finding in the generated PDF report includes the slm_score and slm_confirmed fields

What the model does NOT do

It does not detect names of persons (handled by spaCy NER on filenames)
It does not make decisions — it produces scores for human review
It does not process or store the scanned content beyond the inference session

Explainability

Each detected entity contributes a traceable output:

python

{
  "type": "EMAIL",
  "value": "j***.d*****@h******.fr",   # masked in report
  "slm_score": 0.923,
  "slm_confirmed": True
}

The confidence score is derived from the model's softmax output on the token span overlapping the regex-detected value. Scores below 0.55 are surfaced in the report as uncertain rather than hidden.

GDPR / HDS Compliance Notes

Local inference only: no data is sent to external APIs or servers
No persistent storage of scanned content: only metadata (file paths, PII types, scores) are stored
Purpose limitation: the model is used exclusively for compliance auditing
Legal basis: legitimate interest + legal obligation (GDPR Art. 32 — security of processing)

Citation

If you use this model in research or compliance documentation, please cite:

@software{yalen_artemesia_2026,
  author = {Yalen AI},
  title = {Yalen Artemesia — PII Detection Platform},
  year = {2026},
  url = {https://huggingface.co/yalen-ai/xlm_roberta_yalen}
}

Additional Resources

Full audit documentation: docs/MODEL_CARD_AUDIT.md in the Yalen Sentinel Pulse repository
NER pipeline source: Aiko/aiko/pii_detector.py
SLM validation source: Aiko/aiko/slm_validator.py
spaCy NER model: fr_core_news_sm

About Yalen AI

Yalen Sentinel Pulse is an AI-powered platform for PII detection and data privacy protection, developed by the Yalen AI team. It combines regex patterns, ML models, and NER to provide comprehensive sensitive data identification.

🌐 Platform: Yalen Sentinel Pulse
📧 Contact: hello@yalen.ai

License

MIT — free for commercial and research use.