Apps · PROJECT NOTES
OS-Atlas Data
A cross-platform GUI grounding pretraining corpus combining newly captured desktop/web interfaces with reused mobile and web datasets.
What can you do with it?
Train screenshot-to-element grounding using referring instructions and normalized bounding boxes from Windows, macOS, Linux, Android and web interfaces. Inspect the acquisition and mixture of desktop, mobile and web sources.
What should you know before using it?
This release is static grounding training data, not a live environment, a task-completion benchmark or a collection of independent interaction trajectories. Its large reported count measures GUI elements. Several source components reuse AMEX, Rico/UIBert, SeeClick and other datasets, so they must not be counted as new original tasks. Training requires prompt wrapping and the documented coordinate convention. The repository and dataset card declare Apache-2.0; original component and screenshot rights remain relevant. Associated OS-Atlas models and ScreenSpot-v2 are not counted here.
What is verified here?
Runtime: Public JSON annotations and multipart image archives; records use img_filename, instruction and normalized [left, top, right, bottom] boxes. Platform tags refer to collected interfaces, not software installed or reset by this artifact.
License note: The OS-Atlas repository license and OS-Atlas-data card declare Apache-2.0. This declaration should not be read as replacing the terms of every reused upstream component or third-party interface asset.
This entry was reviewed against original project material. We have not independently run the software or reproduced its results. AI-development claims are author-reported; check upstream for current versions, setup requirements, and licenses.
- Resource type
- Dataset
- Provenance
- Development provenance not established
- Reviewed
- Platforms
- Windows, macOS, Linux, Android, Web