HVSBench

HVSBench HVSBench is a benchmark for evaluating how well multimodal large language models align with human perceptual behavior. It covers human visual system tasks across prominence, subitizing, prioritizing, free-viewing, and searching. This Hugging Face release packages the divided 10% test subset described in the paper as parquet shards with embedded image bytes. It contains 8,657 question-answer examples and 7,507 unique raw images referenced through the image column. The… See the full description on the dataset page:

Source: Hugging Face Hub (garrying/HVSBench). Metadata imported from the dataset’s Hub tags.