redmadrobot-rnd/pii_train

Russian PII NER Training Dataset

Dataset Description

This is the training corpus for PII (Personally Identifiable Information) detection and Named Entity Recognition (NER) on Russian-language text. It targets guardrail and anonymization pipelines that must reliably find personal data (names, addresses, contacts) and Russian identity-document numbers (passport, SNILS, INN, OMS, etc.) in text.

The corpus combines real, manually annotated examples from production logs — where all real personal data has been replaced with synthetic equivalents — with synthetic document-style texts and hand-filtered hard negatives. It is token-level annotated in the standard BIO scheme across 21 entity types.

Dataset Sources

• Evaluation set: redmadrobot-rnd/pii_benchmark — the held-out test split for this corpus (2,841 sentences, identical schema, disjoint rows)
• Model trained on this corpus: redmadrobot-rnd/rubert-base-pii-ner — the NER component of the pii-guard anonymization pipeline
• Code: github.com/redmadrobot-rnd/pii-guard

Relation to the benchmark

This dataset is the train counterpart of redmadrobot-rnd/pii_benchmark (2,841 evaluation sentences): identical label set, annotation scheme, and three-column schema, disjoint rows.

Content Categories

The corpus is annotated across 21 entity types, grouped into four families.

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.

Training Sample Composition

The dataset contains 17,137 sentences with 39,687 annotated entity spans.

Table 1. Entity spans per type

Entity	Count	Entity	Count	Entity	Count
FIRST_NAME	3,876	URL	2,090	BIRTH_CERTIFICATE	1,590
LAST_NAME	3,361	COUNTRY	1,999	MILITARY_ID	1,442
CITY	2,669	HOUSE	1,906	INN	1,163
PASSPORT	2,540	EMAIL	1,755	CREDIT_CARD	989
MIDDLE_NAME	2,344	DISTRICT	1,696	IP_ADDRESS	939
STREET	2,190	DRIVER_LICENSE	1,688	SNILS	924
PHONE	2,134	REGION	1,650	OMS	742

Table 2. Spans grouped into coarse categories (the schema used for cross-model comparison on pii_benchmark)

Coarse category	Fine types	Spans
LOCATION	COUNTRY, REGION, DISTRICT, CITY, STREET, HOUSE	12,110
PERSON	FIRST_NAME, LAST_NAME, MIDDLE_NAME	9,581
RU_DOC_ID	SNILS, OMS, MILITARY_ID, BIRTH_CERTIFICATE	4,698
PASSPORT	PASSPORT	2,540
PHONE	PHONE	2,134
URL	URL	2,090
EMAIL	EMAIL	1,755
DRIVER_LICENSE	DRIVER_LICENSE	1,688
INN	INN	1,163
CREDIT_CARD	CREDIT_CARD	989
IP_ADDRESS	IP_ADDRESS	939

Data Structure

Column	Type	Description
text	string	The sentence as a single string
tokens	string	JSON list of word tokens
ner_tags	string	JSON list of BIO tags, one per token

The 21 entity types appear in the BIO scheme as B-<TYPE> / I-<TYPE> (plus O), e.g. B-FIRST_NAME, I-STREET, B-SNILS — 43 labels in total.

Curation Rationale

General-purpose Russian NER corpora cover person and location names but not the Russian identity-document taxonomy (passport, SNILS, INN, OMS, driver licence, military ID, birth certificate) that an anonymization guardrail has to redact, and they contain no negative examples for numeric strings that merely look like document identifiers. This corpus was built to fill both gaps, with a fine-grained address hierarchy so that redaction can be selective rather than blanket.

Source Data

Table 3. Source families

Family	Rows	Spans	Description
Annotated logs	9,940	25,042	Real user queries, messages, and machine/log-style text, pseudonymized, then BIO-labeled token-by-token and manually verified.
Synthetic documents	6,437	14,355	Generated templates per Russian document type, with formatting and casing variation, to stress structured-identifier detection. Labels are emitted by the generator alongside the text, so spans are exact by construction.
Hard negatives	760	290	Borderline examples — numbers, codes, and measurements that look like document identifiers but are not — filtered manually to confirm no span should be labeled.

Table 4. Synthetic document templates, rows per type

Template	Rows	Template	Rows
mixed documents (free-form)	977	passport	496
mixed documents	898	INN	491
short texts	671	SNILS	475
birth certificate	510	credit card	468
driver licence	502	OMS policy	448
military ID	501

Provenance is documented here but is not shipped as a column — the rows form a single flat table, exactly like the benchmark.

Personal and Sensitive Information

The corpus is about personal data but is not intended to contain any. All real personal data in the log-sourced rows was replaced with synthetic equivalents prior to annotation; document numbers, cards, names, addresses, emails, phones, and IPs are synthetic and do not identify real people. Document identifiers are format-valid (including checksums where applicable) so that detectors can be trained on realistic inputs — do not treat them as issued documents.

Citation

If you use this dataset, please cite this repository.

BibTeX:

@misc{redmadrobot_pii_train,
  title        = {Russian PII NER Training Dataset},
  author       = {Red Mad Robot R\&D},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/redmadrobot-rnd/pii_train}},
  note         = {Companion evaluation set: \url{https://huggingface.co/datasets/redmadrobot-rnd/pii_benchmark}}
}