--- dataset_info: features: - name: video_id dtype: string - name: title dtype: string - name: os dtype: string - name: num_scenes dtype: int64 - name: scene_timestamps_in_sec sequence: float64 - name: screen_bboxes sequence: sequence: int64 - name: ui_element_bboxes sequence: sequence: sequence: float64 - name: raw_actions list: list: - name: box_id dtype: int64 - name: details dtype: string - name: type dtype: string - name: actions list: list: - name: action_type_id dtype: int64 - name: action_type_text dtype: string - name: annot_position sequence: float64 - name: lift sequence: float64 - name: touch sequence: float64 - name: type_text dtype: string - name: video_fps dtype: float64 - name: video_width dtype: int64 - name: video_height dtype: int64 splits: - name: train num_bytes: 69622260 num_examples: 19725 - name: validation num_bytes: 1641036 num_examples: 495 - name: test num_bytes: 565401 num_examples: 100 - name: test_unseen_os num_bytes: 169823 num_examples: 50 download_size: 18085770 dataset_size: 71998520 configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* - split: test path: data/test-* - split: test_unseen_os path: data/test_unseen_os-* --- [Paper](https://arxiv.org/abs/2505.12632) | [Code](https://github.com/runamu/monday) | [Dataset](https://huggingface.co/datasets/runamu/MONDAY) | [Project](https://monday-dataset.github.io) # Dataset Card for MONDAY _MONDAY_ (Mobile OS Navigation Task Dataset for Agents from YouTube) is a cross-platform mobile navigation dataset for training vision-language models. This dataset contains - **20K** curated list of videos of mobile navigation tasks from YouTube, including Android and iOS devices. - **333K** detected scenes, each representing a temporally segmented step within a mobile navigation task. - **313K** identified actions, including touch, scroll, hardware, typing, long press, multi touch and zoom. Please visit our [project page](https://monday-dataset.github.io/) for more details. ## Data Fields - **video_id (str)**: Unique identifier for the video. - **title (str)**: Title of the video. - **os (str)**: Operating system of the mobile device used in the video. - **num_scenes (int)**: Number of detected scenes in the video. - **scene_timestamps_in_sec (list)**: A list of timestamps of the detected scenes in seconds. The list has a length of `num_scenes`. - **screen_bboxes (list)**: A list of bounding boxes for the detected phone screen in each scene, given as (left, top, right, bottom) pixel coordinates. The list has a length of `num_scenes`. - **ui_element_bboxes (list)**: A list of bounding boxes for the detected user interface (UI) elements in each scene, given as (left, top, right, bottom) coordinates normalized to the [0, 1] range. The list has a length of `num_scenes - 1`. ```python # example ui_element_bboxes = [ [ui_bbox1_scene1, ui_bbox2_scene1, ...], # UI elements in scene 1 [ui_bbox1_scene2, ui_bbox2_scene2, ...], # UI elements in scene 2 ... ] ``` - **raw_actions (list)**: A list of raw actions identified from the video for each scene. The list has a length of `num_scenes - 1`. Multiple actions can be annotated within a single scene, and all are considered valid. Each element is a list of actions annotated in that scene, with each action represented as a dictionary containing the following keys: - **box_id (int)**: The index of the UI element's bounding box (from `ui_element_bboxes[scene_id]`) associated with the action. If the action does not correspond to any UI element, the value is -1. - **details (str)**: A detailed description of the action, either automatically generated or manually annotated during the identification process. - **type (str)**: A text label describing the action type. Possible values include `"touch"`, `"scroll"`, `"hardware"`, `"typing"`, `"long press"`, `"multi touch"` and `"zoom"`, listed in order of frequency in the dataset. ```python # example raw_actions = [ [ {"box_id": 0, "details": "...", "type": "touch"}, # First action in scene 1 {"box_id": 1, "details": "...", "type": "touch"}, # Second action in scene 1 ], [ {"box_id": -1, "details": "...", "type": "typing"}, # First action in scene 2 ], ... ] ``` Note: The `box_id` is -1 for actions that do not correspond to any UI element. - **actions (list)**: A list of actions in each scene, processed for mobile navigation agent training and evaluation. The list has a length of `num_scenes - 1`. Multiple actions can be annotated within a single scene, and all are considered valid. Each element is a list of actions annotated in that scene, with each action represented as a dictionary containing the following keys: - **action_type_id (int)**: An integer identifier for the action type, based on the action type sets used in [SeeClick](https://github.com/njucckevin/SeeClick/blob/main/agent_tasks/action_type.py) and [AitW](https://github.com/google-research/google-research/blob/master/android_in_the_wild/action_type.py). - **action_type_text (str)**: A text label describing the action type. Possible values include `"click"`, `"scroll down"`, `"press home"`, `"type"`, `"scroll up"`, `"other hardware"`, `"scroll left"`, `"zoom or multi-touch"`, `"press power"`, `"scroll right"`, and `"press back"`, listed in order of frequency in the dataset. - **annot_position (array)**: A flattened array of bounding box coordinates for detected UI elements, formatted as (top, left, height, width), normalized to the [0, 1] range, and rounded to three decimal places. If applicable, the length of this array is `4 * num_ui_elements` per scene; otherwise, it is an empty list. - **lift (array)**: Lift coordinates in (x, y) format, normalized to the [0, 1] range and rounded to three decimal places. If not applicable, the value is (-1, -1). - **touch (array)**: Touch coordinates in (x, y) format, normalized to the [0, 1] range and rounded to three decimal places. If not applicable, the value is (-1, -1). - **type_text (str)**: The entered text, if the action type is `"type"`; otherwise, this is an empty string. ```python # example actions = [ [ {"action_type_id": 4, "action_type_text": "click", "annot_position": annot_position, "lift": lift_point_action1, "touch": touch_point_action1, "type_text": ""}, # First action in scene 1 {"action_type_id": 4, "action_type_text": "click", "annot_position": annot_position, "lift": lift_point_action2, "touch": touch_point_action2, "type_text": ""}, # Second action in scene 1 ], [ {"action_type_id": 3, "action_type_text": "type", "annot_position": [], "lift": [-1, -1], "touch": [-1, -1], "type_text": "..."}, # First action in scene 2 ], ... ] ``` Note: The data format of `actions` is derived from [SeeClick](https://github.com/njucckevin/SeeClick/blob/main/agent_tasks/readme_agent.md) and [AitW](https://github.com/google-research/google-research/tree/master/android_in_the_wild#dataset-format). - **video_fps (float)**: Frames per second of the video. _This value must be preserved when downloading the video to ensure consistency with `scene_timestamps_in_sec`._ - **video_width (int)**: Width of the video in pixels. _This value must be preserved when downloading the video to ensure consistency with `screen_bboxes`._ - **video_height (int)**: Height of the video in pixels. _This value must be preserved when downloading the video to ensure consistency with `screen_bboxes`._ ## Citation ```bibtex @inproceedings{jang2025_monday, title={{Scalable Video-to-Dataset Generation for Cross-Platform Mobile Agents}}, author={Jang, Yunseok and Song, Yeda and Sohn, Sungryull and Logeswaran, Lajanugen and Luo, Tiange and Kim, Dong-Ki and Bae, Kyunghoon and Lee, Honglak}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year={2025} } ````