kierandesmond/spanish-gdpr-pii-ner-v9

Spanish GDPR PII NER — v9

Token-classification (NER) model that detects GDPR / LOPDGDD-regulated personal data in Spanish text, including standard identifiers and GDPR Art. 9 special-category data (health, genetic, biometric, ethnic, political, religious, union membership, sexual orientation, criminal records).

• —Base model: `kierandesmond/spanish-gdpr-pii-ner-v8` (continued fine-tuning, NOT trained from scratch)
• —Architecture: XLMRobertaForTokenClassification (xlm-roberta-base, ~277M params)
• —Labels: 57 BIO tags / 28 entity types (schema identical to v6/v8 — preserved verbatim)
• —Max length: 256

What's new in v9

v9 was driven by a large, statistically-meaningful held-out evaluation set (7,304 examples, 200–336 per entity, with hard negatives) — replacing the tiny n=1–6 battery used for v6/v8 that masked real weaknesses. Re-scoring v8 on this honest eval exposed gaps (e.g. SIPCARD 0.5%, ETHNICORIGIN 40%, GENETIC_DATA 70%) that the old battery rated as 80–100%.

v9 adds targeted real + synthetic training data for the weak entities and hard negatives for the false-positive traps, while preserving locked entities.

Evaluation (large held-out eval set, span-level pass-rate)

Every one of the 28 entity types scores ≥ 90%. Biggest gains vs v8: SIPCARD 0.5%→100%, ETHNICORIGIN 40%→100%, GENETICDATA 70%→94.6%, CREDITCARD 71%→100%, BANKACCOUNT 80%→100%, POLITICALOPINION 81%→97.6%, UNIONMEMBERSHIP 82%→100%, BIOMETRICDATA 82%→100%, USERNAME 85%→100%, SEXUALORIENTATION 88%→100%, PERSONNAME 90%→100%. (One minor trade-off: HEALTH_DATA 99.4%→97.0%, still well above bar.)

Confusion guards are clean: SEXUALORIENTATION↔RELIGIOUSBELIEF = 0 confusion; POLITICALOPINION↔UNIONMEMBERSHIP = 0 confusion.

Full per-entity numbers, comparison CSV, and confusion matrix: `kierandesmond/spanish-gdpr-pii-ner-v9-validation`.

Training data sources (v9)

Real corpora (parquet-native, no script execution):

• —HEALTH_DATA — `BSC-NLP4BIA/bsc-bio-distemist-ner` (ENFERMEDAD, train split)
• —GENETIC_DATA — `PlanTL-GOB-ES/pharmaconer` (PROTEINAS, train split)
• —Anchor / locked entities (PERSONNAME, ADDRESSES, DATEOFBIRTH, EMAIL, PHONE_ES, NUSS) — `bigbio/meddocan` (char-offset → BIO, train split) for catastrophic-forgetting protection

Synthetic (special categories have no labeled Spanish corpus) — register-varied templates (legal / informal / clinical), contrastive minimal pairs for confusion guards, hard negatives, and regex-generated structured identifiers (DNI/NIE/NIF/CIF/IBAN/SIP/credit-card/etc.). Anti-leakage: synthetic eval uses template families disjoint from training; real-corpus eval uses validation/test splits never seen in training.

Training configuration

• —Continued from v8 weights; LR 1e-5, 3 epochs, effective batch 32 (16×2), warmup 0.1, weightdecay 0.01, bf16, maxlength 256.
• —Final token-level seqeval on held-out eval: precision 0.965, recall 0.976, F1 0.971.

Entity types (28)

PERSONNAME, EMAIL, PHONEES, ADDRESSES, DATEOFBIRTH, IPADDRESS, URL, USERNAME, DNI, NIE, NIF, CIF, NUSS, IBANES, SIPCARD, PASSPORTES, VEHICLEPLATEES, CREDITCARD, BANKACCOUNT, HEALTHDATA, GENETICDATA, BIOMETRICDATA, ETHNICORIGIN, POLITICALOPINION, RELIGIOUSBELIEF, UNIONMEMBERSHIP, SEXUALORIENTATION, CRIMINALRECORD.

Usage

from transformers import pipeline
ner = pipeline("token-classification", model="kierandesmond/spanish-gdpr-pii-ner-v9",
               aggregation_strategy="simple")
ner("El trabajador está afiliado a CCOO y fue diagnosticado de diabetes mellitus tipo 2 .")

Limitations & honest caveats

• —Special-category recall partly depends on phrasing patterns. Although v9 uses register-varied, template-family-disjoint synthetic data (so eval measures generalization, not memorization), production free-text will be more varied than any synthetic set. Treat special-category detection as assistive (flag-for-review), not fully autonomous redaction, until validated on real production documents.
• —No real labeled corpus exists for Art. 9 special categories in Spanish — those entities are synthetic-trained, grounded in researched real-world phrasings.
• —The eval set, while large (200–336/label), is built from public corpora + synthetic generation; a production-grade acceptance test would add human-labeled documents from the target domain.
• —This model is a privacy/compliance aid and does not guarantee complete PII removal.