Turkish clinical NER for PII masking

A fine-tune of akdeniz27/bert-base-turkish-cased-ner that finds person names in Turkish clinical text (lab reports, discharge summaries, consult notes, nursing notes, ward handover lists), including lower / upper-case names and names with case suffixes ("Öztürk'ün", "Koç'tan"). It is the name detector of the PII-masking gateway in secure-clinical-llm, where it runs together with checksum-validated rules for ID numbers, phones, e-mails and IBANs and a fail-closed privacy gate.

Weights are stored in float16 (220 MB). The label set is unchanged from the base model (O, B/I-PER, B/I-ORG, B/I-LOC); only PER was supervised during fine-tuning.

Usage

from transformers import pipeline

ner = pipeline("token-classification", model="yagmurtuncer/bert-base-turkish-clinical-ner",
               aggregation_strategy="simple")
ner("Hasta ayşe yılmaz'ın kontrolü Dr. Mehmet Kaya tarafından yapıldı.")

Inside secure-clinical-llm: HFNer("yagmurtuncer/bert-base-turkish-clinical-ner"), or set SCLLM_NER_MODEL to this id.

Training

Data: 2,000 synthetic documents from four template families (lab headers, discharge summaries, consult notes, referral letters) with names from one half of the name lists; 15% of documents fully upper- or lower-cased. Generator: src/scllm/pii/synthetic.py.
1 epoch, AdamW, lr 3e-5, batch 8, max length 192, CPU (about 15 minutes). Script: scripts/train_ner.py. training.json in this repo holds the exact settings and loss log.

Evaluation

Test set	Base model	This model
1,000 synthetic documents, held-out template families and names: name recall	0.988	1.000
same set: documents with ≥1 identifier left unmasked (full gateway)	2.0 %	0 % (95% upper bound 0.38 %)
WikiANN Turkish test, 1,000 real sentences: person-name F1	0.724	0.705

The held-out set checks generalization to document formats and names the model never saw; WikiANN checks that fine-tuning on synthetic text did not break name detection on real text (a small precision loss, no collapse).

Limitations

Trained and evaluated on synthetic documents written by the same author. Performance on real hospital reports is not measured; validate on your own de-identified data before relying on it.
Only person names are tuned. Use rule-based, checksum-validated detectors for ID numbers, phone numbers, e-mail addresses, IBANs and dates, and keep a recall-first second check before data leaves a trusted zone.
Not a medical device and not a compliance guarantee (KVKK / GDPR / HIPAA).

加载中…

Turkish clinical NER for PII masking

Weights are stored in float16 (220 MB). The label set is unchanged from the base model (O, B/I-PER, B/I-ORG, B/I-LOC); only PER was supervised during fine-tuning.

Usage

from transformers import pipeline ner = pipeline("token-classification", model="yagmurtuncer/bert-base-turkish-clinical-ner", aggregation_strategy="simple") ner("Hasta ayşe yılmaz'ın kontrolü Dr. Mehmet Kaya tarafından yapıldı.")

Inside secure-clinical-llm: HFNer("yagmurtuncer/bert-base-turkish-clinical-ner"), or set SCLLM_NER_MODEL to this id.

Training

Data: 2,000 synthetic documents from four template families (lab headers, discharge summaries, consult notes, referral letters) with names from one half of the name lists; 15% of documents fully upper- or lower-cased. Generator: src/scllm/pii/synthetic.py.

1 epoch, AdamW, lr 3e-5, batch 8, max length 192, CPU (about 15 minutes). Script: scripts/train_ner.py. training.json in this repo holds the exact settings and loss log.

Evaluation

Test set	Base model	This model
1,000 synthetic documents, held-out template families and names: name recall	0.988	1.000
same set: documents with ≥1 identifier left unmasked (full gateway)	2.0 %	0 % (95% upper bound 0.38 %)
WikiANN Turkish test, 1,000 real sentences: person-name F1	0.724	0.705

Limitations

Trained and evaluated on synthetic documents written by the same author. Performance on real hospital reports is not measured; validate on your own de-identified data before relying on it.

Only person names are tuned. Use rule-based, checksum-validated detectors for ID numbers, phone numbers, e-mail addresses, IBANs and dates, and keep a recall-first second check before data leaves a trusted zone.

Not a medical device and not a compliance guarantee (KVKK / GDPR / HIPAA).