Turkish PII (Personally Identifiable Information) detection model for KVKK compliance. Fine-tuned from dbmdz/bert-base-turkish-cased for token classification.

Model Description

This model detects unstructured PII entities in Turkish text that cannot be caught by deterministic regex patterns. It is designed as Tier 2 of the kvkk-pii hybrid pipeline.

Base model: dbmdz/bert-base-turkish-cased (110M parameters)
Task: Token classification (BIO tagging)
Labels: PERSON, ADDRESS, HEALTH_DATA, KVKK_SPECIAL
Training data: 14,913 synthetic samples
Name generation reference: Nisanyan Adlar Sozlugu

Labels

Label	Description	KVKK Category
PERSON	Person names	Kimlik (Md. 3)
ADDRESS	Physical addresses	Iletisim (Md. 3)
HEALTH_DATA	Diagnoses, medications, lab results	Ozel Nitelikli (Md. 6)
KVKK_SPECIAL	Religion, ethnicity, political views	Ozel Nitelikli (Md. 6)

Usage

With kvkk-pii (recommended)

python

from kvkk_pii import PIIAnalyzer

analyzer = PIIAnalyzer(ner_model_path="ersanbil/kvkk-pii-ner")
results = analyzer.analyze("Hasta: Ahmet Yilmaz, Tani: E11.9 - Tip 2 diabetes mellitus")

Standalone with transformers

python

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("ersanbil/kvkk-pii-ner")
model = AutoModelForTokenClassification.from_pretrained("ersanbil/kvkk-pii-ner")
model.eval()

text = "Hasta Adi: LEVENT DUMAN"
encoding = tokenizer(text, return_tensors="pt", return_offsets_mapping=True)

with torch.no_grad():
    outputs = model(**{k: v for k, v in encoding.items() if k != "offset_mapping"})
    predictions = outputs.logits.argmax(dim=-1)[0]

for token_id, pred in zip(encoding["input_ids"][0], predictions):
    label = model.config.id2label[pred.item()]
    if label != "O":
        print(f"{tokenizer.decode([token_id])}: {label}")

Performance

Entity	Precision	Recall	F1
PERSON	1.00	1.00	1.00
ADDRESS	1.00	1.00	1.00
HEALTH_DATA	0.99	1.00	0.99
Overall			99.86%

5-Fold Cross-Validation: Mean F1 = 98.86% +/- 0.19%

Limitations

KVKK_SPECIAL category has limited training coverage
City names may be tagged as PERSON (acceptable from a PII perspective)
Free-text health data recall is lower than structured forms
Inference runs on CPU for cross-device consistency (110M model is fast enough)

Training

Fine-tuned with HuggingFace Transformers
10 epochs, batch size 64, learning rate 3e-5, cosine scheduler
Early stopping with patience 3

License

Apache License 2.0

Author

Ersan Bilik — ersan.ai

Citation

bibtex

@software{kvkk_pii_ner,
  author = {Bilik, Ersan},
  title = {kvkk-pii-ner: Turkish PII Detection Model for KVKK Compliance},
  year = {2026},
  url = {https://huggingface.co/ersanbil/kvkk-pii-ner}
}