Human Kirkpatrick Annotation Dataset for Teaching Innovations in Medical Education (2000-2025)

Human-annotated modified-Kirkpatrick outcome-level labels for a 1,200-record diagnostic subset of a medical-education literature corpus (2000-2025) covering twelve predefined teaching innovations. Each record was independently labelled by two annotators, blind to each other, after a pilot-calibration round; all 481 disagreements were resolved by an independent adjudicator. Contents. A single UTF-8 CSV (1,200 rows) giving, per record, the innovation category, the pre/post-COVID stratum, each annotator's twelve-class label, the adjudicator's label and free-text rationale (for adjudicated records), the disagreement type, and the final consensus label. A codebook documents the twelve-class label space (six Kirkpatrick levels L1-L4b plus six non-applicable categories), the decision rules, the disagreement-type taxonomy, and the annotation/adjudication protocol. Data quality. Two-annotator agreement was 59.9% raw; Cohen's kappa 0.541 (twelve-class) and 0.612 (seven-class collapsed); weighted kappa 0.757 on the ordinal Kirkpatrick L1-L4b subset. What is NOT included, and why. Copyrighted abstract text and bibliographic metadata are not redistributed; the records were drawn from a corpus that includes subscription-database (Scopus, Web of Science) content. To respect subscription-database licensing, records are keyed by an opaque within-dataset identifier and are therefore not linkable to specific external publications. Raw Scopus and Web of Science exports are excluded under Elsevier and Clarivate licence restrictions. Analysis code and derived/aggregate tables are out of scope for this data-only deposit.