JABD — Job Advertisements Bias Dataset

JABD is a token-level, human-annotated corpus for studying linguistic bias in job advertisements. 14,960 sentences drawn from 8,993 real postings are annotated under a four-level taxonomy separating explicit from implicit bias across five sociodemographic groups (gender, ethnicity, age, religion, disability), with 34,598 flagged spans over 12 leaf classes. To our knowledge JABD is the first token-level bias resource for job advertisements. Existing resources annotate whole postings or whole sentences, which shows that an advertisement may be biased but not which expression makes it so. This deposit contains the annotation layer only. It does not redistribute the full text of the source advertisements. The upstream corpus, EMSCAD (University of the Aegean, 2012–2014), is no longer reachable at its official host, so its redistribution terms cannot be verified; the source advertisements also carry employer contact details. A script is included to re-attach the advertisement text for users who obtain EMSCAD themselves. Two exploratory classes. Old Coded and Young Coded have the lowest inter-annotator agreement and the thinnest multiply-validated support of any class. Benchmarks should be reported over the ten core classes; see the datasheet. Not for decisions about people. JABD detects properties of text. It must not be used to screen, rank, score or reject candidates. See the use restrictions in LICENSE. Funding. ANID Fondecyt grant 1241462. The annotation campaign was approved by the Institutional Review Board of the Pontificia Universidad Católica de Chile; annotators were recruited through Prolific between June and July 2025, participated on an informed and voluntary basis, and were compensated according to that platform's protocol. Changes in v2.1. The datasheet is complete: funding, ethics approval, annotator compensation and campaign dates are filled in, and the DOI is recorded in the README and CITATION.cff. The data files are byte-identical to v2.0.