benchmark-pii-recall

PII extraction recall is measured against synthetic gold spans using exact span and grapheme-level scoring to evaluate if models detect sensitive entities reliably before deployment, while reporting aggregate metrics without exposing identifier text.

Recall benchmarking can enforce release gating by requiring zero misses for critical identifiers and applying recall floors to reject model releases that fall below required privacy thresholds.

PII recall evaluation supports comparing quantized and full-precision models, tracking device-specific slices, and evaluating backend, language, and script variations for privacy-preserving clinical text processing.

Privacy-safe evaluation compares exact-span and grapheme-level recall while reporting aggregate metrics only, preventing identifier surfaces from appearing in evaluation output.

Synthetic recall benchmarking fixtures are built with exact offsets, canonical labels, boundary cases, and multilingual or multiscript coverage to test PII detection across diverse text formats.

Aggregate-only evidence limits debugging by restricting exposure to identifier text, requiring normalized labels and hard negatives to ensure PII recall scoring accuracy without compromising privacy.