pii-synthetic-basque-v2

A fully synthetic dataset of 6,000 sentences containing personally identifiable information (PII), in Basque, Spanish and French, covering the whole Basque Country (Araba, Bizkaia, Gipuzkoa, Nafarroa and Iparralde / Pyrénées-Atlantiques). It is the v2 of pii-synthetic-basque: unlike v1, realistic values (name frequencies, streets, towns, postcodes) are drawn from public aggregate statistics and open map data, never from records of real individuals.

It was created as experience replay data for fine-tuning multilingual PII detection models (e.g. GLiNER2-based ones) on Basque while preserving their original PII types — see the model tknika/gliner2-PII-basque and its v2, tknika/gliner2-PII-basque-v2. Its intended use is educational: anonymizing student texts (essays, forum posts, school records) before sharing or analyzing them.

Data sources

All values that look realistic come from public, aggregate sources:

• INE (Instituto Nacional de Estadística, Spain) — given-name and surname frequencies per province (public statistics from its name/surname widgets). Used to weight how common each name is in Araba, Bizkaia, Gipuzkoa and Nafarroa. Candidates from EODA plus the most common Spanish surnames were queried and kept when frequent enough nationally (≥ 300 bearers for names, ≥ 200 as first surname for surnames): 757 given names and 890 surnames made it into the final tables.
• INSEE (France) — Fichier des prénoms, département 64 (Pyrénées-Atlantiques): given-name frequencies for Iparralde (600 names, including compound French names such as Jean-Pierre).
• EODA (Euskaltzaindia's Basque onomastics database) — the universe of Basque given names and surnames (≈ 5,000 names, ≈ 10,000 surnames), used as the candidate pool for the INE queries.
• OpenStreetMap (© OpenStreetMap contributors) — street names, towns and postcodes, extracted from Geofabrik dumps (País Vasco, Navarra, Aquitaine) via address points: 17,029 streets in 824 towns, always with the town and postcode the street actually has.

Everything else (emails, phone numbers, national IDs, IBANs, card numbers, dates) is randomly generated by the included script.

Guarantees

• No personal data. Person names are built by sampling a given name and a surname independently from two marginal frequency distributions (INE / INSEE public statistics). No name pair is taken from any record of a real individual. Common names may coincidentally form a plausible real identity — that is an unavoidable property of aggregate statistics, not personal data.
• Streets, towns and postcodes are public map data, always combined coherently (a street appears with the town and postcode it actually has in OSM).
• Numeric values (Spanish DNI, French NIR, ES/FR IBANs, card numbers, phone numbers) are randomly generated with valid checksums (mod-23 DNI letters, mod-97 NIR/IBAN check digits, Luhn-valid card numbers) so that they look realistic to models, but they are not taken from any source of real data.
• Caveat: a randomly generated phone number, email address or NIR may coincidentally coincide with a real one, as with any random generator; no value in this dataset is associated with, or derived from, any individual.

Territorial and linguistic profile

Sentences are sampled by territory (Araba 5%, Bizkaia 25%, Gipuzkoa 25%, Nafarroa 20%, Iparralde 25%) and language is drawn per territory: Basque/Spanish in the southern territories (70/30 in the EAE, 35/65 in Nafarroa) and Basque/French in Iparralde (35/65). Contact details follow each territory: Spanish phone formats (+34 / 0034 / 6XX XXX XXX, 94X landlines) vs. French ones (+33 / 0033 / 06-07 mobiles, 05 59 landlines), Spanish DNI vs. French NIR, ES vs. FR IBANs, and territory-typical email domains (euskaltel.eus, ehu.eus… vs. orange.fr, free.fr…).

Format

One JSON object per line, in two files (pii-sintetikoa-v2-train.jsonl and pii-sintetikoa-v2-eval.jsonl):

{"text": "Kaixo, Mikel Etxeberria naiz eta Amorebietan bizi naiz. Nire helbidea Santa Luzia auzoa 105 da, 48340 postakodearekin...",
 "entities": {"person name": ["Mikel Etxeberria"], "city": ["Amorebietan"], "address": ["Santa Luzia auzoa 105"], "zip code": ["48340"], ...},
 "spans": [[7, 24, "person name"], [40, 49, "city"], [67, 87, "address"], [92, 97, "zip code"], ...]}

entities maps GLiNER-style entity types to the list of mentions found in text (each mention is an exact substring in the surface form it appears in, declension suffixes included — e.g. in «Amorebietan bizi naiz» the mention is Amorebietan, not Amorebieta); spans provides the character offsets of each mention. The dataset is provided as two splits: pii-sintetikoa-v2-train.jsonl (5,000 examples, for fine-tuning / experience replay) and pii-sintetikoa-v2-eval.jsonl (1,000 examples, for evaluation). Splits are generated independently (seeds 42 / 2025) and are disjoint by construction:

• no sentence appears in both splits;
• all person-identifying values (names, emails, usernames, personal URLs) are disjoint across splits, so evaluation cannot be inflated by memorization;
• negative sentences (15% of each split, no entities) are split in disjoint halves per language;
• streets, towns, organizations and numeric values may appear in both splits, mirroring real-world distributions.

Territory and language proportions are sampled identically in both splits (see the territorial and linguistic profile below).

Each line is a JSON object with text, entities (GLiNER-style type → list of exact-mention strings) and spans (character offsets [start, end, type] for span-exact evaluation). Types covered: person name, email, phone number, national id, bank account number, credit card number, date of birth, date, address, city, location, zip code, age, organization, plus three types specific to educational contexts (following the taxonomy of the Kaggle PII Detection, Removal from Educational Data competition): user name (forum/LMS handles), personal url (blogs and personal profiles) and student id (school/university student numbers, distinct from national IDs). The location type is reserved for provinces and historical territories (Araba, Bizkaia, Gipuzkoa, Nafarroa, Lapurdi, Nafarroa Beherea, Zuberoa); town names are always typed as city.

Notes

• Basque sentences include real declension suffixes on town, street and organization names (e.g. «Amorebietan bizi naiz», «Donostiako kalea», «El Correoren sisteman»), which makes the data useful for testing agglutinative handling; the generator derives them with standard Basque rules, and the gold mentions carry the declined form exactly as it appears in the sentence.
• Several templates place a national ID (DNI/NIR) and a student ID in the same sentence with distinct phrasing (e.g. «NAN 34780993W eta ikasle-zenbakia 865776B»), as hard negatives for identifier disambiguation.
• Every mention in both splits (20,139 in total) is token-aligned with the WhitespaceTokenSplitter used by the GLiNER2 training pipeline, so the data passes strict token-alignment validators.
• ~15% of the sentences are PII-free negatives, and every sentence lists 2–4 entity types that do not appear with empty lists, as hard negatives for the entity types.
• The generator script is included in generator/ for full transparency and reproducibility (deterministic; seeds 42 / 2025 for the two splits).

Citation

If you use this dataset, please cite:

@misc{tknika2026piisyntheticbasquev2,
  title  = {pii-synthetic-basque-v2: A Synthetic PII Dataset for Basque Experience Replay with Educational Entity Types},
  author = {{TKNIKA} and Ezpeleta Mendikute, Xabier},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tknika/pii-synthetic-basque-v2}
}

The synthetic-data methodology follows the GLiNER2-PII paper, which you may also want to cite:

@misc{zaratiana2026gliner2piimultilingualmodelpersonally,
  title  = {GLiNER2-PII: A Multilingual Model for Personally Identifiable Information Extraction},
  author = {Zaratiana, Urchade and Lewis, Ash and Hurn-Maloney, George},
  year   = {2026},
  eprint = {2605.09973},
  archivePrefix = {arXiv},
  url   = {https://arxiv.org/abs/2605.09973}
}