neondijital/neonredact-tr-web

NeonRedact-TR Web

Turkish PII detection that runs in the browser. 68 MB quantized ONNX, no server, no upload — the text never leaves the user's machine.

Live demo: neondijital.com/kisisel-veri-bul

Open PII models are trained on English and extended to a handful of other languages. Turkish is in none of them. This is the gap this model fills.

Usage

import { pipeline } from '@huggingface/transformers';

const nlp = await pipeline(
  'token-classification',
  'neondijital/neonredact-tr-web',
  { dtype: 'q8' }
);

const metin = "Müvekkilim Ayşe Öztürk'ün T.C. kimlik numarası 85592807206.";
const sonuc = await nlp(metin, { aggregation_strategy: 'simple' });

Returns BIO-tagged spans. Note that start and end may be absent depending on the runtime version; if so, locate the returned word in the source text yourself.

Labels

16 entity types, BIO scheme (33 tags including O).

Label	Meaning
KISI_ADI	person name
ADRES	address, full or partial
TC_KIMLIK	Turkish national ID
VERGI_NO	tax identification number
IBAN	Turkish IBAN
KART_NO	payment card number
CEP_TELEFON	mobile phone
SABIT_TELEFON	landline
EPOSTA	email address
PLAKA	vehicle licence plate
DOGUM_TARIHI	date of birth
KURUM_ADI	organisation, court, hospital
MESLEK	occupation
AKRABALIK	kinship reference
SAGLIK_DURUMU	health condition
TARAF_SIFATI	party role (client, defendant, patient)

Why Turkish needs its own model

Suffixes stay outside the span. Turkish attaches case endings to proper nouns with an apostrophe. A model that swallows the suffix produces spans that break redaction: replace Ayşe Öztürk'ün wholesale and the sentence loses its grammar. This model returns Ayşe Öztürk and leaves 'ün alone. The same holds for the pronominal buffer in institution names — Ege Üniversitesi out of Ege Üniversitesi'nden.

Turkey-specific identifiers. T.C. national ID, tax number, Turkish IBAN and licence plates have no equivalent in EU-oriented datasets, so EU-oriented models do not look for them.

People named by role. In Turkish legal and clinical writing the subject is often a role rather than a name — müvekkilim, davalı, hastam.

Files

File	Size	Notes
onnx/model.onnx	270 MB	fp32, exact
onnx/model_quantized.onnx	68 MB	int8, the browser default (dtype: 'q8')

Quantization agreement with the fp32 model, measured over 9,814 tokens of held-out data: 99.36% of token predictions identical.

Training

Fine-tuned from dbmdz/distilbert-base-turkish-cased on neondijital/neonredact-tr — 10,000 synthetic records, 36 document templates. Four epochs on a single T4, about two minutes.

Subword tokens inherit their word's label (B- becoming I-), so multi-piece identifiers come back whole rather than fragmented.

This architecture replaced an earlier GLiNER-based attempt. GLiNER was more flexible but 611 MB at its smallest, and its ONNX graph resisted quantization — int8 shifted the logits by 21.7 in testing. A plain token classifier is both an order of magnitude smaller and, on the held-out probes, more accurate: it detects bare landline numbers, which three rounds of GLiNER training never managed, and does not absorb inflectional suffixes into occupation spans.

The larger GLiNER model is still available at neondijital/neonredact-tr-model for server-side use.

Known limitations

• Trained on synthetic, template-generated text. Real documents are messier.
• No OCR noise, no layout artefacts, no dialectal spelling.
• TARAF_SIFATI is the weakest label — it has the fewest training examples.
• Occupation and health-condition vocabularies are limited, not exhaustive.
• Compound and hyphenated surnames are not covered.

Recommended use

Pair it with deterministic checks. Checksum-verifiable identifiers — T.C. ID, tax number, IBAN, card number — should be validated by their own algorithms, which are exact where a model only guesses. Use the model for what rules cannot reach: names, addresses, kinship, occupations, roles. The demo linked above runs both layers and lets the rule layer win on overlap.

Detection is not compliance. Removing direct identifiers does not make a document anonymous under KVKK or GDPR — a residual combination of attributes (birth year, district, occupation) can still single out one person. Keep a human in the loop before anything leaves the building.

Citation

@misc{neonredact_tr_web_2026,
  title  = {NeonRedact-TR Web: In-Browser Turkish PII Detection},
  author = {Kömürcü, Aykan},
  year   = {2026},
  url    = {https://huggingface.co/neondijital/neonredact-tr-web}
}

Built by NeonDijital.