MapTab

MapTab is a multimodal benchmark for map understanding, map-based question answering, and constrained multi-criteria route planning over heterogeneous topological graphs. This complete package contains MetroMap and TravelMap queries, referenced map images, edge/vertex tables, prompt templates, normalized Parquet files, lossless raw source records, a local/Hub loader, validation tooling, and an upload interface.

• Paper: arXiv:2602.18600
• Code: Ziqiao-Shang/MapTab
• Project page: MapTab Leaderboard

Dataset configurations

Configuration	Splits	Rows	Contents
metromap_qa	test	2,080	13 QA tasks × 160 maps
travelmap_qa	test	2,184	13 QA tasks × 168 maps
metromap_planning	train/test	76,800 / 19,200	12 planning query variants
travelmap_planning	train/test	80,640 / 20,160	12 planning query variants

The normalized planning configs omit the small demo_*.json helper files and avoid duplicating data/all, because all is exactly the union represented by train and test. The lossless source query records remain under raw/; demo helper files are intentionally excluded from the public package and inference interface.

Repository layout

Maptab_hug/
├── README.md                 # canonical Hugging Face dataset card
├── dataset_card.md           # standalone copy of the dataset card
├── data/                     # normalized Parquet configurations
├── assets/
│   ├── metromap/{images,tabulars}/
│   └── travelmap/{images,tabulars}/
├── raw/
│   ├── metromap/{data,qa_data,prompts}/
│   └── travelmap/{data,qa_data,prompts}/
├── manifest.json             # counts, sizes, and generated shards
├── maptab_loader.py          # resolves image/table paths after download
└── tools/{build_dataset,validate,upload}.py

Test-only snapshot for the inference repository

The official inference code and this package use the same directory contract. Download only the test queries and their assets with:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="szq-nju/MapTab",
    repo_type="dataset",
    revision="main",
    local_dir="./data",
    allow_patterns=[
        "README.md",
        "dataset_card.md",
        "manifest.json",
        "maptab_loader.py",
        "requirements.txt",
        "data/metromap_qa/test-*.parquet",
        "data/travelmap_qa/test-*.parquet",
        "data/metromap_planning/test-*.parquet",
        "data/travelmap_planning/test-*.parquet",
        "assets/metromap/**",
        "assets/travelmap/**",
        "raw/metromap/qa_data/**",
        "raw/travelmap/qa_data/**",
        "raw/metromap/data/test_set/**",
        "raw/travelmap/data/test_set/**",
        "raw/metromap/prompts/**",
        "raw/travelmap/prompts/**",
    ],
)

The same operation is exposed by the code repository as maptab-infer download --subset test --data-root ./data.

Native Hugging Face loading

The Parquet configurations work directly with datasets:

from datasets import load_dataset

ds = load_dataset("szq-nju/MapTab", "metromap_qa", split="test")
print(ds[0])

The native loader exposes image, edge_table, and vertex_table as repository-relative paths. To resolve those paths and optionally decode images, use the included helper after cloning or downloading this repository:

from maptab_loader import load_maptab

ds = load_maptab(
    "metromap_planning",
    repo_id="szq-nju/MapTab",
    split="test",
    decode_images=True,
)
print(ds[0]["image"])

Local use is also supported:

from maptab_loader import load_maptab

ds = load_maptab("travelmap_qa", local_dir=".", split="test")

Normalized schema

Every row uses one stable schema across domains and task families:

• id, domain, task_family, task, split
• question, answer, routes
• image, edge_table, vertex_table
• country, location, origin, destination
• map_difficulty, query_difficulty, modality, question_type
• weights, route/transfer counts, aggregate route attributes
• source_file, source_index, and raw_json

raw_json preserves every original field exactly, including source-specific fields that are not part of the normalized schema.

Build and validate

The checked-in package can be rebuilt from an authoritative MapTab workspace:

python tools/build_dataset.py --source /path/to/MapTab --output .
python tools/validate.py --root .

Validation checks row counts, stable IDs, Parquet readability, and every referenced image/table path.

Upload

Authenticate with hf auth login, then run:

python tools/upload.py --repo-id szq-nju/MapTab --root .

The upload interface uses HfApi.upload_folder, supports private repositories and revisions, and can be rerun after interruption.

Dataset creation and preprocessing

Normalized files are generated deterministically from:

• <domain>/qa_data/*.json for QA;
• <domain>/data/training_set/*.json and test_set/*.json for planning;
• <domain>/images and <domain>/tabulars for referenced assets.

QA CSV files are retained in raw/ but are not duplicated as normalized rows because they mirror the JSON records. Planning CSV modality variants are inference-time views over the same query records and tabular assets rather than separate labels.

Audited source-path repair

The original TravelMap task-13 QA records reference 168 nonexistent *_vertex2.json files. Every task-13 label was checked against the released *_vertex.json table and matched exactly (168/168). Normalized Parquet rows therefore reference the real *_vertex.json assets. The original path remains unchanged in raw/ and raw_json, and the repair is recorded in manifest.json.

Limitations

Map diagrams vary in resolution, visual style, language, and topology. Strict route accuracy can mark a response wrong after a single omitted station. The dataset contains repeated maps across task variants by design. Users should report the exact configuration, split, prompt, model route, decoding settings, and image-processing budget.

License

The source snapshot and upstream GitHub repository did not expose a standalone license file when this package was assembled. Accordingly, the Hub metadata uses license: other. Before public upload, the dataset owner should replace this section and the metadata value with the authoritative dataset license and confirm redistribution rights for all map images and tables.

Citation

@article{shang2026maptab,
  title={MapTab: A Diagnostic Benchmark for Long-Horizon Multi-Criteria Multimodal Reasoning on Heterogeneous Topological Graphs},
  author={Shang, Ziqiao and Ge, Lingyue and Xu, Zian and Cheng, Zi-Jian and Tian, Shi-Yu and Huang, Zhenyu and Fu, Wenbo and Wu, Weiming and Chen, Yang and Zhang, Xiangwen and Hu, Yulan and Liu, Bin and Guo, Lan-Zhe},
  journal={arXiv preprint arXiv:2602.18600},
  year={2026}
}