InciText — Incident-Centric Text Dataset for Attribute Extraction

InciText v2 is a cleaned public release of the Incident-Centric Text Dataset for person-attribute extraction from heterogeneous incident-related and generated text. The release contains 572 source-text documents: a 558-document consolidated core and the 14-document Challenge14 extension. The consolidated core includes 10 WLPD Synthetic descriptions, 13 press releases, 5 incident/dispatch documents, 30 Generated documents, and 500 Generated_flash documents. Challenge14 adds 13 incident/dispatch-style documents and one press release. The Annotation/ directory contains canonical report metadata and person-level annotations for the 558-document core and Challenge14, explicit paper-cohort ID lists, a minimal loading example, and auxiliary date/time/location records. The release also includes de-identified annotations for Facebook21, an external evaluation cohort of 21 randomly selected public missing-person notices from Missing People in America. The raw Facebook source posts are not redistributed because they contain personally identifiable information, including information concerning minors. This v2 release does not carry forward the Purdue Exponent collection, FemmIR/PostgreSQL-derived exports, Generated_flash_2, or other legacy/intermediate material included in InciText v1. Those materials remain available through the archived v1 release. InciText v2 is released for research and academic use. Users are responsible for following applicable privacy, ethical, institutional, and source-specific requirements.See the included README.md files for release structure, cohort definitions, and loading instructions.