| --- |
| language: |
| - en |
| tags: |
| - computer-use |
| pretty_name: Real-World Computer Use Agent Training Data |
| --- |
| |
| # Pango Sample: Real-World Computer Use Agent Training Data |
|
|
| **Pango** represents **P**roductivity **A**pplications with **N**atural **G**UI **O**bservations and trajectories. |
|
|
| ## Dataset Description |
|
|
| This dataset contains authentic computer interaction data collected from users performing real work tasks in productivity applications. The data was collected through [Pango](https://pango.so), a crowdsourced platform where users are compensated for contributing their natural computer interactions during actual work sessions. |
|
|
|  |
|
|
| ## Motivation |
|
|
| Current Computer Use Agent (CUA) training datasets face several limitations: |
| - **Scale constraints**: Existing datasets like Mind2Web (2,350 tasks), GUI-World (12,000 videos), and OSWorld (369 tasks) provide limited coverage |
| - **Artificial contexts**: Most demonstrations are scripted rather than authentic work sessions |
| - **Distribution gaps**: Performance drops significantly when agents encounter interfaces outside their training distribution |
| - **Missing error patterns**: Academic datasets typically exclude "failed" interactions, removing important recovery behaviors |
|
|
| This dataset addresses these limitations by capturing real users performing genuine work tasks, providing natural interaction patterns, error recovery sequences, and diverse problem-solving approaches. |
|
|
| ## Data Collection Methodology |
|
|
| Data is collected through a Chrome extension that records user interactions during structured "quests" in target applications: |
| - **Applications**: Google Sheets, Google Slides, Figma, Canva (more coming soon) |
| - **User base**: Global contributor network across 180+ countries |
| - **Task context**: Authentic work sessions (financial analysis, presentation creation, design work, etc.) |
| - **Compensation**: Users are paid based on session length and data quality |
|
|
| ## Dataset Structure |
|
|
| Each record contains: |
| - `id`: Unique session identifier |
| - `video_url`: Screen recording of the interaction session |
| - `input_metadata`: Structured JSON containing granular interaction events |
| - `input_metadata_preview_url`: Link to interactive data viewer for the `input_metadata` |
| - `task_description`: User-provided description of what they were doing |
| - `quest_type`: Application category (Sheets, Slides, Figma, Canva) |
| - `profession`: User's professional background |
| - `synthetically_generated_instruction`: Synthetically generated task instruction for training purposes. Represents the context of the full task. |
| - `synthetically_generated_thought_metadata`: (Beta) Synthetically generated thoughts for each user step. Represents the thought of the current step. Available by request. |
| - `artifact_history_preview_url`: Link to the full history of the data artifact. Currently only Canva data. Useful for code generation in addition to pure CUA. |
|
|
| ### Input Metadata Schema |
|
|
| The `input_metadata` field contains timestamped interaction events with the following structure: |
|
|
| ```json |
| { |
| "relative_timestamp_ms": 1028, |
| "type": "click", |
| "x": 186.0, |
| "y": 62.445, |
| "button": "button_left", |
| "screenshot_url": "https://...", |
| "click_count": 1 |
| } |
| ``` |
|
|
|
|
| **Key fields:** |
| - `relative_timestamp_ms`: Milliseconds since session start |
| - `type`: Event type (click, input, key_press, mouseover_start, mouseover_end, drag_start, drag_end, scroll) |
| - `x,y`: Screen coordinates (normalized for display resolution) |
| - `screenshot_url`: URL to corresponding interface screenshot |
| - `text`: Text content for input events - only available on input events |
| - `key_codes`: Keyboard key identifier (DOM KeyboardEvent codes) - only available on key_press events |
| |
| ### Previewing the Data |
| |
| While the `video_url` is provided for completeness, it is often clunky to interact with `webm` files and it's impractical to leverage these for most use cases. |
|
|
| The `input_metadata` is a much more convenient way to interact with the data, however it is not as easy to visualize. To solve this, we built an interactive preview tool, which takes the `input_metadata` and renders it frame-by-frame with corresponding actions. |
|
|
| This can be found [here](https://admin.pango.so/actions-preview?url=https://pango-service-production.s3.us-east-1.amazonaws.com/finalized/bc50d07a-d11a-4c9c-9c5a-e5fdb6e85b5f/actions_v2.json) and it takes any valid `input_metadata` URL as input. |
|
|
|  |
|
|
| ## Thought Metadata (Beta) |
|
|
| An additional field, `synthetically_generated_thought_metadata`, is included to provide synthetically generated thoughts for each user step. This field is designed to enhance the dataset's utility for training reasoning VLMs like [UI-TARS 1.5](https://huggingface.co/ByteDance-Seed/UI-TARS-1.5-7B). It is not to be confused with `synthetically_generated_instruction`, which is the context of the full task. This field is available by request. |
|
|
| **Step Generation and Aggregation** |
|
|
| To create `thought_metadata`, we begin with the `input_metadata` where each row represents an individual user action. As a first stage, we aggregate actions into steps, where each step represents either a single action or a collection of actions. |
|
|
| **Batch Processing Strategy** |
|
|
| We partition the steps into batches with the following parameters: |
|
|
| $$ |
| \alpha = 7, \quad \beta = 15, \quad \gamma = 15 |
| $$ |
|
|
| where: |
| - \\(\alpha\\) (**pre_window_size**): Number of steps preceding the target step used for context |
| - \\(\beta\\) (**post_window_size**): Number of subsequent steps used for context |
| - \\(\gamma\\) (**batch_size**): Interval between target steps for thought generation. i.e. if \\(\gamma = 15\\), then for every 15 steps, we generate a thought. |
| |
| The \\(\gamma = 15\\) prevents overlapping thoughts and ensures that thoughts generated in earlier batches are considered completed when generating subsequent thoughts. |
| |
| **LLM Usage** |
| |
| The processed step batches are fed to GPT-4o with its vision API, using high image detail settings. Each thought generation process consumes approximately 30,000 input tokens. |
| |
| ## Quality Assurance |
| |
| Data quality is maintained through: |
| - Automated filtering of invalid interactions and privacy-sensitive content |
| - Quality scoring based on task coherence and completion patterns |
| - Compensation algorithms that reward genuine engagement |
| - Differential privacy techniques to prevent individual behavior reconstruction |
| |
| ## Use Cases |
| |
| This dataset is designed for: |
| - Training computer use agents on authentic interaction patterns |
| - Studying human-computer interaction behaviors across diverse populations |
| - Developing more robust GUI automation systems |
| - Research on temporal reasoning and error recovery in sequential decision-making |
| |
| ## Ethical Considerations |
| |
| - All users provide informed consent for data collection and redistribution |
| - Privacy-sensitive content is automatically filtered |
| - Compensation ensures fair value exchange for user contributions |
| - Data collection follows ethical guidelines for crowdsourced research |
| |
| ## Data Scale and Growth |
| |
| The dataset is continuously growing through ongoing collection: |
| - Planned: Scaling to 100,000+ hours over 2025 |
| |
| ## Citation |
| |
| If you use this dataset in your research, please cite: |
| |
| ```bibtex |
| @dataset{pango2025, |
| title={Pango: Real-World Computer Use Agent Training Data}, |
| author={Chakra Labs}, |
| year={2025}, |
| url={https://huggingface.co/datasets/chakra-labs/pango} |
| } |
| ``` |
| |
| ## Contact |
| |
| For access to the full dataset or collaboration opportunities, please contact [Chakra Labs](mailto:nirmal@chakra-labs.com). |
| |
| |
