SHIELD Dataset

SHIELD (Synthetic Human-annotated Identifier-replaced Entries for Learning and De-identification) is a diverse clinical note dataset designed for the development and evaluation of de-identification systems for electronic health records. The dataset contains 1,381 de-identified clinical notes with 10,229 gold-standard PHI (Protected Health Information) spans across 9 entity categories, sourced from the Stanford Medicine STARR-OMOP clinical data warehouse. Notes were selected via set-cover diversity sampling across demographic and document-type strata, ensuring broad representation of clinical language variation. All notes underwent human-in-the-loop annotation with expert adjudication to establish high-quality gold-standard PHI boundaries.SHIELD is intended for the evaluation of PII detection and de-identification systems on clinical text. Every note has been de-identified using a cryptographic surrogate-replacement pipeline, which replaces each PHI span with a type-appropriate surrogate that preserves the linguistic structure and readability of the original text. The dataset is provided in Parquet format with train/validation/test splits and is compatible with standard NLP evaluation frameworks.