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.
Fix T2 question types in task overview table
Update dataset card (new configs and supervision conditions)
Upload t4 images (query condition)
Upload t3 images (query condition)
Upload t2 images (query condition)
Upload t1 images (query condition)
Remove legacy direct/oracle-named files (superseded by renamed conditions)
Upload benchmark jsonl (anonymized, renamed conditions)
Fix loading example with actual repo id
Upload dataset card
Upload benchmark images
Upload benchmark jsonl (anonymized)
initial commit