| --- |
| license: apache-2.0 |
| task_categories: |
| - robotics |
| - image-to-text |
| tags: |
| - computer-use |
| - gui-agent |
| - osworld |
| - trajectories |
| - reinforcement-learning |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-*.parquet |
| --- |
| |
| # Computer Use Trajectories |
|
|
| Successful computer-use agent trajectories collected on [OSWorld](https://os-world.github.io/) tasks. |
|
|
| ## Dataset Details |
|
|
| - **Rows**: 160 (one per task trajectory) |
| - **Steps**: 1,378 total across all trajectories (avg ~8.6 steps/task) |
| - **Agent**: Gemini 3 Flash Preview with linearized accessibility-tree grounding |
| - **Score filter**: Only trajectories with score = 1.0 (fully successful) |
|
|
| ### Domains |
|
|
| | Domain | Tasks | Description | |
| |--------|-------|-------------| |
| | chrome | 21 | Web browsing tasks in Google Chrome | |
| | gimp | 15 | Image editing in GIMP | |
| | libreoffice_calc | 17 | Spreadsheet tasks in LibreOffice Calc | |
| | libreoffice_impress | 20 | Presentation tasks in LibreOffice Impress | |
| | libreoffice_writer | 11 | Document editing in LibreOffice Writer | |
| | multi_apps | 21 | Cross-application workflows | |
| | os | 15 | Operating system / desktop tasks | |
| | thunderbird | 10 | Email tasks in Thunderbird | |
| | vlc | 8 | Media player tasks in VLC | |
| | vs_code | 22 | Code editing in VS Code | |
| |
| ## Schema |
| |
| Each row is a complete task trajectory. Per-step data is stored as parallel lists (index `i` across all list columns gives step `i`). |
| |
| | Column | Type | Description | |
| |--------|------|-------------| |
| | `task_id` | string | UUID of the task | |
| | `domain` | string | Application domain (e.g. "chrome", "gimp") | |
| | `instruction` | string | Natural-language task instruction | |
| | `score` | float | Final task score (all 1.0 since filtered to successful) | |
| | `num_steps` | int | Number of steps in this trajectory | |
| | `actions` | list[string] | Executed pyautogui actions (Python code or JSON) | |
| | `responses` | list[string] | Full LLM responses including `<think>` reasoning | |
| | `screenshots` | list[Image] | PNG screenshots of the desktop at each step | |
| | `accessibility_trees` | list[string] | Linearized accessibility tree at each step | |
| | `exe_statuses` | list[string] | Execution status per step ("success" / "error") | |
| | `exe_outputs` | list[string] | Stdout from each action execution | |
| | `exe_errors` | list[string] | Stderr from each action execution | |
| | `rewards` | list[float] | Per-step reward (0 for intermediate steps) | |
| | `dones` | list[bool] | Whether the episode ended at each step | |
| | `recording_path` | string | Relative path to MP4 screen recording | |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("markov-ai/computer-use") |
| |
| # Each row is a full trajectory |
| traj = ds["train"][0] |
| print(traj["instruction"]) |
| print(f"{traj['num_steps']} steps") |
| |
| # Iterate through steps |
| for i in range(traj["num_steps"]): |
| screenshot = traj["screenshots"][i] |
| action = traj["actions"][i] |
| print(f"Step {i+1}: {action[:80]}...") |
| |
| # Filter by domain |
| chrome_tasks = ds["train"].filter(lambda x: x["domain"] == "chrome") |
| ``` |
|
|
| ## Recordings |
|
|
| MP4 screen recordings are stored in `recordings/{domain}/{task_id}.mp4`. |
| These are full-episode videos capturing the agent's desktop interaction. |
|
|
| ## Source |
|
|
| Collected using the [ComputerRL](https://github.com/anthropics/ComputerRL) evaluation framework on OSWorld virtual machines. |
|
|
