NeonRedact-TR: Turkish PII Detection Dataset

A synthetic, span-annotated dataset for detecting personally identifiable information (PII) in Turkish text.

Existing open PII models are trained on English and extend to a handful of other languages. Turkish is not among them. This dataset exists to close that gap.

Why Turkish needs its own dataset

Agglutinative suffixes. Turkish attaches case suffixes to proper nouns with an apostrophe: Ahmet'in, Konak'tan, Ege Üniversitesi'nde. A model that marks the suffix as part of the entity produces spans that break downstream redaction. In this dataset the suffix is always outside the entity span:

text:   Zeynep Yıldırım'ın dosyası incelenmektedir.
entity: [0, 16] KISI_ADI -> "Zeynep Yıldırım"

Vowel harmony, consonant assimilation (Konak'ta vs Ahmet'te) and the pronominal buffer -n- after third-person possessive endings (Belediyesi'nde, not Belediyesi'de) are all generated correctly.

Turkey-specific identifiers. T.C. national ID, tax number (VKN), Turkish IBAN and licence plates have no equivalent in EU-oriented datasets.

Contextual party references. In Turkish legal and clinical writing a person is often named by role rather than by name (müvekkilim, davalı, hastam). These carry identifying weight and are annotated separately.

Labels

16 labels in two groups.

Format-verifiable (checksum-validated at generation time):

Label	Description
TC_KIMLIK	Turkish national ID (11 digits, checksum valid)
VERGI_NO	Tax identification number (10 digits, checksum valid)
IBAN	Turkish IBAN (26 chars, mod-97 valid)
KART_NO	Payment card number (Luhn valid)
CEP_TELEFON	Mobile phone number
SABIT_TELEFON	Landline number
EPOSTA	Email address
PLAKA	Vehicle licence plate

Contextual (require sentence-level understanding):

Label	Description
KISI_ADI	Person name
ADRES	Address, full or partial
DOGUM_TARIHI	Date of birth
KURUM_ADI	Organisation, court, hospital, company
MESLEK	Occupation
AKRABALIK	Kinship reference
SAGLIK_DURUMU	Health condition
TARAF_SIFATI	Party role (client, defendant, patient)

Format

JSONL, one record per line:

{
  "text": "Müvekkilim Ayşe Öztürk'ün T.C. kimlik numarası 85592807206.",
  "entities": [
    {"start": 11, "end": 22, "label": "KISI_ADI", "text": "Ayşe Öztürk"},
    {"start": 46, "end": 57, "label": "TC_KIMLIK", "text": "85592807206"}
  ]
}

Offsets are character-level over the raw string.

Generation

10,000 records built from 36 document templates: court petitions, service of process, clinical notes, invoices, tenancy agreements, job applications, corporate correspondence and everyday messaging, in both formal and informal register.

Deliberately included hard cases: identifiers preceded by their own label (TCKN 10000000146 -> the label is outside the span), IBANs written with and without spaces, occupations carrying inflectional suffixes (eczacıymış), honorifics (Sn., Ayşe Hanım), and age/sex descriptors in clinical notes that must not be tagged. Landline and mobile numbers appear in the same sentence so the two can be told apart, and document reference numbers (Fatura No: 2026/4471) sit alongside real IBANs so the model learns which long number is which.

No real person's data is present. Names are drawn from common Turkish given names and surnames and randomly paired. Identifiers are generated to be format-valid but random. Addresses are composed from real place names with fabricated street numbers.

Every record is verified before release: span offsets match the text exactly, no overlapping spans, and every checksum-bearing identifier passes its own algorithm. Current run: 0 errors across all checks.

Roughly 8% of records are adversarial negatives containing numbers that resemble identifiers but are not PII (case file numbers, order codes, monetary amounts), to suppress false positives.

Limitations

• Template-generated. Sentence structures are more regular than real documents. Fine-tuning on this alone will not fully generalise to free-form prose. Validation scores computed on a held-out split of this dataset are optimistic; evaluate on your own documents.
• Not OCR-derived, so it contains no scanning noise or layout artefacts.
• Occupation and health condition pools are limited and not exhaustive.
• Dialectal and informal spelling variation is under-represented.
• Compound and hyphenated surnames are not covered.

Intended use

Fine-tuning token classification models for PII detection in Turkish, so that documents can be redacted or pseudonymised before they are sent to any third-party service.

Detection is not a compliance guarantee. Removing direct identifiers does not by itself make a document anonymous under KVKK or GDPR: a residual combination of attributes can still single out an individual. Keep a human in the loop.

Citation

@misc{neonredact_tr_2026,
  title  = {NeonRedact-TR: A Synthetic Turkish PII Detection Dataset},
  author = {Kömürcü, Aykan},
  year   = {2026},
  url    = {https://huggingface.co/datasets/neondijital/neonredact-tr}
}

Built by NeonDijital.