mapspatial/map-spatial-benchmark

Map-based Spatial Reasoning Benchmark A multi-view map-based spatial reasoning benchmark. Each row is one multiple-choice question instance over a registered map image; models must answer with a single option letter. Four tasks (T1–T4), four base-map views, and controlled evidence conditions (direct / query / oracle) and world perturbations (transform / world layers) allow fine-grained analysis of spatial reasoning robustness. Task overview Task Question… See the full description on the dataset page: https://huggingface.co/datasets/mapspatial/map-spatial-benchmark.