| --- |
| pretty_name: PIE2F-Visuals |
| license: cc0-1.0 |
| task_categories: |
| - image-feature-extraction |
| - visual-question-answering |
| tags: |
| - computer-use |
| - gui |
| - temporal |
| - privacy-filtered |
| --- |
| |
| # PIE2F-Visuals |
|
|
| PIE2F-Visuals is a dataset for studying long-horizon computer-use tasks. It is built from |
| more than 600 hours of real-world workflow recordings across browsers, development tools, |
| research work, and other desktop environments. |
|
|
| This gated early preview contains **2 privacy-filtered visual states** and |
| **1 temporal transitions** sampled across **1 |
| source sessions**. The complete corpus is still being processed and validated. |
|
|
| Each state may include safe OCR, UI elements, a 768-dimensional visual embedding, and a |
| controlled semantic label. Rapid selection uses a small number of full-timeline anchors so |
| adjacent near-identical screens do not dominate the preview or waste OCR computation. |
|
|
| The source recordings come from a permissively licensed public corpus. Versioned release |
| metadata records source and model revisions, processing configuration, validation checks, |
| and artifact hashes. |
|
|
| ## Preview status |
|
|
| This is a bounded preview, not the complete release. Access requests are reviewed manually. |
| Automated privacy filtering reduces risk but does not guarantee that every sensitive detail |
| has been removed. Users should perform their own review before training or evaluation. |
|
|
| Developed by Parergon Labs. |
|
|
