Datasets:
| language: | |
| - en | |
| - zh | |
| size_categories: | |
| - 1K<n<10K | |
| task_categories: | |
| - image-text-to-text | |
| pretty_name: OmniMapBench | |
| # OmniMapBench | |
| OmniMapBench is a map-centric visual question answering benchmark. This release contains 2,096 QA samples over 1,603 unique map images. | |
| - **Paper:** [OmniMapBench: Benchmarking Visual-Centric Reasoning on Diverse Map Documents](https://huggingface.co/papers/2607.09068) | |
| - **Repository:** [SIGMME/OmniMapBench](https://github.com/SIGMME/OmniMapBench) | |
| ## Dataset Structure | |
| The repository is organized as: | |
| ```text | |
| README.md | |
| dataset_infos.json | |
| metadata.jsonl | |
| data/ | |
| images/ | |
| ``` | |
| Each row in `metadata.jsonl` contains: | |
| - `image_path`: relative image path inside this dataset repo. | |
| - `id`: original QA id. | |
| - `question`: question text. | |
| - `choices`: answer candidates. | |
| - `answer`: ground-truth answer. | |
| - `qa_type`: single_choice, multi_choice, or sort. | |
| - `capability_level`: level_1, level_2, or level_3. | |
| - `language`: en or zh. | |
| - `type`: map category. |