A triplet-based hallucination evaluation dataset for visual question answering. Contains 1226 flattened question-answer instances across 300 images. Each row includes the question, answer, and associated knowledge triplets (both instance-level and image-level) encoded as JSON strings. The triplets capture subject-predicate-object relationships used to probe visual hallucination.
