Held-out evaluation set, scripts, and results for
kierandesmond/spanish-gdpr-pii-ner-v9.
v6/v8 were scored on a tiny n=1–6 "battery" that reported 90%+ but hid real failures.
v9 introduces a large held-out eval set (7,304 examples, 200–336 per entity, with hard
negatives) to get statistically-meaningful per-entity numbers. Re-scoring v8 on it revealed
SIP_CARD 0.5%, ETHNIC_ORIGIN 40%… See the full description on the dataset page:
https://huggingface.co/datasets/kierandesmond/spanish-gdpr-pii-ner-v9-validation.
