modernbert-pii-v0.1

modernbert-pii-v0.1

ModernBERT-base fine-tuned for PII/PHI span detection over a 169-label BIO space derived from a versioned 84-concept ontology.

Training

Scheme-aware loss

The 32 source corpora annotate different label sets. A corpus that annotates 7 concepts tags every email and phone number as O — under plain cross-entropy those become hard negatives and the model unlearns them. Instead each token is trained against the set of labels consistent with its annotation: for O, that is O plus every concept the source dataset does not annotate.

Training datasets (32)

• ai4privacy/legacy-tiers
• ai4privacy/open-pii-500k
• ai4privacy/openpii-1.5m
• alrosait/russian-pii
• bccard/openpii-finance
• bob14/korean-sensitive
• btx24/turkish-pii
• careons/dutch-healthcare-pii
• datafog/test-corpus
• dataiku/kiji
• dialogpii
• e3jsi/multi-pii-ner
• felipe53/name-judgment
• grascco/multi
• grascco/pii-v2
• gretel/pii-finance-multilingual
• gretel/pii-masking-en
• kdpii
• keler/turkish-medical-deid-eval
• kierandesmond/gdpr-special-categories
• lacuna/pii
• localdoc/azerbaijani-pii
• meddocan
• meddoplace
• meddoprof
• nisaefendiolu/code-secrets
• nutrient/docpii
• nvidia/nemotron-pii
• presidio/research-eval
• sidharth/indicphi
• tab/text-anonymization-benchmark
• tonicai/privacy-bench

Known limitations

• Character offsets from the tokenizer include the preceding space; decoded spans must be whitespace-trimmed. Without it, span-exact accuracy falls from ~99.7% to ~5.9% while token-level F1 is unchanged.
• Spans located by string search rather than given by the annotator are marked derived in the corpus; they train but were never scored.
• Coverage is skewed: health identifiers are a small fraction of training mentions relative to network and contact identifiers.