Job classifier mislabeled 1,200 postings, distorting pay figures for remote roles

A developer who publishes a free dataset of job postings with stated pay ranges found that one fifth of the corpus was sitting in an "other" category, signaling a classifier failure rather than a labeling choice. The original keyword-based classifier matched titles in a fixed order, so roles like Sales Engineer and Support Engineer were wrongly filed as engineering, and short keywords like "data" misfired on titles such as Data Center Technician. After restructuring the classifier into prioritized passes, the "other" bucket fell from 1,726 to 680 postings, with about 1,200 postings reassigned mostly to sales, support, and finance families. The correction also changed a previously published finding: the remote-versus-on-site pay gap within role families shrank from about 11.5% to 6.2%, though the direction of the difference held.