Katip Mask Base · tr v2 (katip-mask-base)

BERTurk (dbmdz/bert-base-turkish-cased) fine-tuned to tag person names, addresses and KVKK art. 6 health terms in Turkish legal documents, as the on-device model for Katip Mask. Exported to ONNX (fp32 and int8) for transformers.js.

v2 (13 Nov 2026) adds 17,967 real Turkish sentences (Turkish Wiki NER, person names only) to the synthetic training documents. v1 found every name in Katip's synthetic test sets but only 41% of names in real Turkish sentences; v2 finds 90%.

Results in Katip (rules + this model, as the demo runs it)

Model names scored 0.5 or higher keep their edges (trust 0.5, chosen on dev data).

Set	v2	v1
Blind (80 synthetic docs, 757 values)	757 / 757, names 365 / 365	757 / 757
Held-out (60 synthetic docs, 618 values)	618 / 618	618 / 618
Name-free legal prose	3 false alarms	3
Real Yargıtay prose, anonymised names filled back in (1,643 names)	99.3% found, 96.2% precision	98.6%, 95.9%
Real Wikipedia sentences, Wiki NER test (654 names)	89.8% [87.1–92.1], 89.9% precision	40.7%

About 27 ms per synthetic document and 19 ms per sentence on an Apple-silicon Mac (int8 ONNX, onnxruntime). Method and scripts: https://github.com/fikrifirat/katip/tree/lab/lab

Results (model alone, no rules)

Trained for 3 epochs on 8,000 synthetic documents plus 2,246 pseudo-documents of eight real sentences each, in 28 minutes on an Apple-silicon Mac (MPS).

Set	Names	Addresses	Health terms	Extra spans
Blind (80 docs, 16 styles, 4 new)	364 / 365	80 / 80	25 / 26	0
Held-out (60 docs)	275 / 275	70 / 70	26 / 26	0
Wiki NER dev (1,000 real sentences)	583 / 641			53

Limits. Addresses and health terms are still learned from synthetic documents only. The real-text court set uses real decisions with names filled in by us, because published Turkish decisions are anonymised; consented real case files from pilot firms are the next test.

Files

File	Size
onnx/model_quantized.onnx	110.8 MB	int8, used by the browser demo
onnx/model.onnx	440.3 MB	fp32
model.safetensors		PyTorch weights, for further fine-tuning

The int8 file is 4× smaller than fp32, and every Katip score above was measured on the int8 file, so quantising cost no accuracy. Side-by-side size, speed and accuracy for all Katip models: https://github.com/fikrifirat/katip#size-speed-and-accuracy-side-by-side

Labels are word-level BIO on each word's first sub-token; use the word split in config.json (katip.word_regex) as Katip's src/ner.js does.

Base model: BERTurk (dbmdz/bert-base-turkish-cased) by Stefan Schweter / dbmdz, MIT licence.

Licence: CC BY-SA 4.0 (see LICENSE). v2 is trained partly on Turkish Wiki NER (Altinok, 2023; CC BY-SA 4.0, https://github.com/turkish-nlp-suite/Turkish-Wiki-NER-Dataset), so the model is shared under the same licence: credit Katip and the dataset, and share adaptations under CC BY-SA 4.0. v1 (synthetic data only) was released under CC BY-NC 4.0. Katip's code is AGPL-3.0; for a commercial licence for the code or support, open an issue at https://github.com/fikrifirat/katip.

Citation

@misc{firat2026katipmaskbase,
  title  = {Katip Mask Base (tr): a Turkish personal-data tagger fine-tuned from BERTurk},
  author = {F{\i}rat, Fikri},
  year   = {2026},
  url    = {https://huggingface.co/f1k/katip-mask-base},
  note   = {Fine-tuned from BERTurk (Schweter, 2020)}
}

Please also cite BERTurk (Stefan Schweter, https://doi.org/10.5281/zenodo.3770924) and Turkish Wiki NER (Altinok, 2023).