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 type	Description
T1	direction / angular_order / egocentric_side / dual_anchor_direction	Cardinal direction, clockwise angular ordering, and egocentric side between colored point markers
T2	composite_euclidean_distance / composite_network_distance / nearest_point / directional_nearest_point	Comparing metric or network distances between marked points, and identifying the nearest point
T3	segment_building_count (subtasks: route, seg)	Counting first-row buildings along a directed route segment
T4	waypoint_ordering / route_validity / detour_waypoint_shortest / detour_waypoint_traversability	Waypoint ordering, route validity, and detour planning under road closures

Supervision conditions

Each task is released under one or more supervision conditions, which differ only in the visual evidence overlaid on the registered map images — the questions, options and answers are identical across conditions of the same task.

File	condition	Available for	Visual evidence
*_no_visual_supervision.jsonl	direct	T1–T4	None — the bare base map, no task annotations
*_query_supervision.jsonl	query	T1–T4	Query-dependent scaffolding that supports answering the question (e.g., route segments, measured rays, partial building-footprint scaffolds)
*_oracle_supervision.jsonl	oracle	T3–T4	Complete oracle annotations of the ground-truth solution

Note: for T4 detour_waypoint_traversability, the query condition uses the fully annotated road-status / waypoint-map pair, since the weak scaffold does not cover this question type.

Sample counts (per condition)

T1/T2 provide 2 conditions each; T3/T4 provide 3. All conditions of a task share the same instance set. Total: 77,800 rows (40,000 for T1/T2 + 13,800 for T3 + 24,000 for T4).

T1 — 10,000 per condition

Base question set = 400 instances per view. 7 geometric transform variants (rot90, rot180, rot270, mirror_h, mirror_h_rot90/180/270) replicate the base set.

view	composition	rows
sat	base 400 + 7 transforms x 400	3,200
wprd01	base 400 + 7 transforms x 400	3,200
blank	base 400 + 7 transforms x 400	3,200
webrd04	base 400 only	400
total		10,000

T2 — 10,000 per condition

Same structure as T1 (400 base instances, same view/transform coverage): 3,200 x 3 + 400 = 10,000.

T3 — 4,600 per condition

Base question set = 200 instances per view; world layer has 2 variants (intervention_001, sham_001) with 200 each. The blank view is not used for T3: the task is building-footprint recognition, which is only meaningful on imagery that actually shows buildings (satellite / map tiles).

view	composition	rows
sat	base 200 + 7 transforms x 200 + 2 world x 200	2,000
wprd01	base 200 + 7 transforms x 200 + 2 world x 200	2,000
webrd04	base 200 + 2 world x 200	600
total		4,600

T4 — 8,000 per condition

Base question set = 400 instances; world layer = 2 variants x 200. The blank view is not used for T4: the task is road/waypoint recognition on the route network, which requires imagery that actually depicts roads (satellite / map tiles).

view	composition	rows
sat	base 400 + 7 transforms x 400 + 2 world x 200	3,600
wprd01	base 400 + 7 transforms x 400 + 2 world x 200	3,600
webrd04	base 400 + 2 world x 200	800
total		8,000

Dimensions

Field	Values	Meaning
view	sat / webrd04 / wprd01 / blank	Base map style: satellite imagery / road map / place-label map / blank background. Note: the webrd04 (road-map) view is only provided for the base layer — its tiles contain textual labels (e.g., Chinese characters), which would become unreadable or mirrored after rotation/flipping, so no transform variants are generated for this view
layer	base / transform / world	Original scene, geometric transform stress test, or world intervention
variant	rot90, rot180, rot270, mirror_h, mirror_h_rot* (transform); intervention_001, sham_001 (world); null (base)	Sub-variant of the layer. Note: transform variants have rotated/flipped map frames — "north is up" applies to the transformed image
condition	direct / query / oracle	Visual evidence condition (see Supervision conditions)

Row schema

{
  "id": "c0019932db15_blank_query",
  "task_id": "T1",
  "layer": "base",
  "variant": null,
  "view": "blank",
  "condition": "query",
  "question_type": "angular_order",
  "question": "From the green point, sweeping clockwise from north, ...",
  "options": ["A. yellow, red", "B. red, yellow"],
  "answer": "B",
  "answer_letter": "B",
  "system_prompt": "You are evaluating spatial reasoning ...",
  "images": ["images/t1/c_06d46000d7/s_429d01__q_e5ebc72207__base__query__blank_query_q4.png"],
  "meta": {
    "case_id": "c_06d46000d7",
    "scheme": "s_429d01",
    "instance_id": "q_e5ebc72207",
    "tile_type": "blank",
    "subtask": null
  }
}

• options are pre-shuffled; the correct letter is answer_letter.
• Evaluation is exact match on the option letter for all rows.
• Multiple-choice prompt format: append Options:\n<options> and instruct the model to output only the letter (see system_prompt).

Image naming

Images live under images/{task}/{case_code}/. Directory segments case_id, scheme and question_dir are anonymized with deterministic short codes (c_*, s_*, q_*); the remaining segments (layer, condition, original file name) are kept verbatim, e.g.

images/t1/c_06d46000d7/s_429d01__q_e5ebc72207__base__query__blank_query_q4.png

The same source image may be referenced by several rows (e.g., shared across transform variants or views); 77,800 rows reference 27,394 unique image files.

Loading

from datasets import load_dataset
from huggingface_hub import snapshot_download
from PIL import Image

# Download the image files once (the jsonl only stores relative image paths)
root = snapshot_download("mapspatial/map-spatial-benchmark", repo_type="dataset")

ds = load_dataset("mapspatial/map-spatial-benchmark", "t1_query_supervision", split="test")
row = ds[0]
img = Image.open(f"{root}/{row['images'][0]}")

Anonymization note

Identifiers of the source map data (case / scheme / question directory) are replaced by deterministic hash codes and are not reversible from this repository.