--- dataset_info: features: - name: images list: image: decode: false - name: inputs dtype: string - name: action dtype: string - name: x10 dtype: int64 - name: y10 dtype: int64 - name: grid5 dtype: int64 - name: grid7 dtype: int64 - name: labels dtype: int64 - name: choices_list list: string - name: metadata dtype: string splits: - name: test_domain num_bytes: 2496471298 num_examples: 2690 - name: test_task num_bytes: 769507248 num_examples: 833 - name: test_website num_bytes: 687395712 num_examples: 634 - name: train num_bytes: 4724111708 num_examples: 5086 download_size: 8672193746 dataset_size: 8677485966 configs: - config_name: default data_files: - split: test_domain path: data/test_domain-* - split: test_task path: data/test_task-* - split: test_website path: data/test_website-* - split: train path: data/train-* license: openrail source_datasets: - osunlp/Multimodal-Mind2Web --- # multimodal-mind2web Multimodal Mind2Web's eligible GUI actions, prepared once for six annotations. Source: osunlp/Multimodal-Mind2Web at 1b4c6a8cf9f77b7a5e0d641959935c80c4a05889. Retains native train, test_task, test_website and test_domain splits, encoded screenshots, stable action IDs and geometry metadata. Inputs contain only the task and past actions. Candidate descriptions omit positive flags. A uniquely marked original target with a valid on-screen box and valid negatives is required. At most 23 negatives are sampled and option order shuffled using the action ID. Canonical labels supervise element selection; action/x10/y10/grid5/grid7 fields provide the derived annotations. No evaluation splits are fabricated. Original Hub card license: OpenRAIL. Original data: [osunlp/Multimodal-Mind2Web](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/](https://github.com/sileod/tasksource/tree/main/scripts/repackage_dataset).