Custom hallucination contrastive dataset — partitions into subtypes of → Hallucination (closed-domain fabrication)

CAA's custom hallucination dataset, generated with GPT-4, does not treat hallucination as one thing to elicit. Following Rawte et al. (2022)'s taxonomy, it structures every contrast pair as either Unprompted Hallucination -- a factually accurate prompt met with distorted or fabricated information -- or Contextually-Triggered Hallucination -- a factually inaccurate prompt, such as a question presupposing that a fictional 'Marauder's Map' influenced the US entering World War I, met with a false narrative built around that false premise rather than a correction (contrastive-activation-addition, §"C Generating custom hallucination dataset", p. 13). Each type gets its own contrast-pair design: unprompted items contrast a valid answer against a fabricated one to a valid question, while contextually-triggered items contrast an answer that flags the premise as false against one that accepts and elaborates it (contrastive-activation-addition, §"C Generating custom hallucination dataset", p. 14). Hallucination's own page can state that it covers both subtypes; only the dataset's page explains that the split comes from an external taxonomy and that each subtype requires a structurally different question to elicit at all.