AiventraLab/

segvqa-review

Human-validation subset of the SegVQA Kvasir-SEG benchmark (segvqa_benchmark_kvasir.json), 100 verified items sampled with seed 42, stratified by category (multi-hop/comparison slightly oversampled).

Category n

morphology 16

multi_hop 18

localization 16

counting 16

comparison 18

existence_detection 16

For each row, mark:

Q — is the question sensible and answerable… See the full description on the dataset page: https://huggingface.co/datasets/AiventraLab/segvqa-review.