Russian PII NER Model

Token-classification model that finds personal data (names, addresses, contacts) and Russian identity-document numbers (passport, SNILS, INN, OMS, driver licence, military ID, birth certificate, bank card) in Russian text. It is the NER component of the pii-guard anonymization pipeline.

Fine-tuned from ai-forever/ruBert-base on 17,137 annotated Russian sentences, with a 21-type / 43-label BIO head.

Model Sources

• Training data: redmadrobot-rnd/pii_train — 17,137 sentences, 39,687 entity spans
• Evaluation data: redmadrobot-rnd/pii_benchmark — 2,841 held-out sentences, disjoint from the training data
• Code: github.com/redmadrobot-rnd/pii-guard

Entity Types

1. Person (names) FIRST_NAME, LAST_NAME, MIDDLE_NAME — given name, surname, and patronymic, including mixed-case, latin-script, and out-of-order spellings.

2. Location / Address COUNTRY, REGION, DISTRICT, CITY, STREET, HOUSE — the full Russian address hierarchy down to the house/building number.

3. Contacts EMAIL, PHONE, URL, IP_ADDRESS — structured contact and network identifiers in many real-world formats.

4. Russian identity-document numbers PASSPORT, INN (taxpayer number), SNILS (insurance account), OMS (medical insurance policy), CREDIT_CARD, DRIVER_LICENSE, MILITARY_ID, BIRTH_CERTIFICATE — document numbers in their canonical and noisy/typo'd forms.

Labels follow the BIO scheme: B-<TYPE> / I-<TYPE> plus O — 43 labels total.

Usage

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="redmadrobot-rnd/rubert-base-pii-ner",
    aggregation_strategy="simple",
)

text = "Иванов Пётр Сергеевич, паспорт 45 11 123456, тел. +7 999 123-45-67"
for ent in ner(text):
    print(f"{ent['entity_group']:16s} {ent['word']!r} {ent['score']:.2f}")

Windowing over long inputs, confidence filtering and merging adjacent fragments of the same entity are the caller's responsibility. The pii-guard pipeline uses max_length=512, a sliding window with stride=128, and drops predictions below min_confidence=0.3.

Merging is not optional: aggregation_strategy="simple" leaves elena@pochta.ru as el + ena@pochta.ru when the model re-emits B- mid-entity. All scores below assume merged fragments.

Evaluation

redmadrobot-rnd/pii_benchmark, 2,841 held-out sentences. Span-level micro-F1, one-to-one greedy matching, gold spans not normalised.

Table 1. 14 categories both systems express — exact matching. Folded: the pipeline emits PERSON where the model emits FIRST_NAME. The pipeline also covers DATE_TIME, BANK_ACCOUNT, BIK and TELEGRAM, which this model has no head for.

On the benchmark's own leaderboard protocol — PERSON + LOCATION, overlap matching — this model scores 94.7 F1, the pipeline 95.0.

Table 2. Per-entity F1, this model alone — exact matching.

IP_ADDRESS is the one weak spot — the model splits on the dots; in pii-guard this type is a regex rule.

Training Data

redmadrobot-rnd/pii_train — 17,137 Russian sentences with 39,687 BIO-annotated entity spans: manually verified production logs (pseudonymized), synthetic document-style texts per Russian document type, and hard negatives (numeric strings that resemble document identifiers but are not). Rows that also occur in pii_benchmark were removed from training, so evaluation stays clean.

Fine-tuned for 10 epochs, lr 3e-5 (linear decay, no warmup), batch size 16, AdamW (weight decay 0.01), fp32, max sequence length 512, seed 42.

Citation

If you use this model, please cite this repository.

BibTeX:

@misc{redmadrobot_rubert_base_pii_ner,
  title        = {Russian PII NER Model},
  author       = {Red Mad Robot R\&D},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/redmadrobot-rnd/rubert-base-pii-ner}},
  note         = {Fine-tuned from ai-forever/ruBert-base}
}