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.
0651
1version https://git-lfs.github.com/spec/v12oid sha256:623007a8cef4e12f4fb9e80acd2c6ae521151ddeeff0f5ad456d7c62ec8b4dd23size 134731784 