OSReward Data Card
Dataset Summary
OSReward contains multimodal examples for outcome judgment of GUI-agent
trajectories across desktop, web, and mobile environments. Each example
provides a user request, selected screenshots, and, depending on the source,
textual action history or a task-specific rubric. The target is an
evidence-based assessment ending with Judge: SUCCESS or Judge: FAIL.
Released Configurations
Supervised fine-tuning
- Exactly 100,000 judge instances from 16 source datasets.
- 359,789 unique screenshots.
- 58,712 SUCCESS labels and 41,288 FAIL labels.
- LLaMAFactory-compatible ShareGPT records.
Reinforcement learning
- Approximately 3K training examples.
- Approximately 200 validation examples.
- Parquet records compatible with the common verl multimodal data schema.
- Screenshots are referenced from the SFT image package and are not embedded in parquet.
Publication Safety Filter
7 records and 29 screenshots in credential-bearing connected components were excluded before publication. This prevents literal API tokens or private-key material found in source GUI tasks from being released.
Intended Use
The data is intended for research on multimodal GUI trajectory evaluation, supervised reward-model training, and rule-reward RL. Users should validate format assumptions, data rights, and suitability for their own deployment.
Limitations
- Labels are outcome judgments derived from heterogeneous GUI tasks and may contain annotation errors or ambiguous cases.
- Platform, task, label, and prompt-style distributions are not uniform.
- Screenshots may contain content rendered by third-party applications or websites.
- The RL split emphasizes difficult and informative cases and should not be treated as an unbiased sample of all GUI interactions.
- Framework and model support for long-context multimodal MoE training is version-sensitive and should be validated in the user's own environment.
No standalone license is specified in this initial release. Users are responsible for determining the rights and restrictions that apply to their use of the records and screenshot content.
WebTrail Supplement
The WebTrail increment contains 6,386 strict-unanimous SFT instances from
6,386 unique web trajectories. Each instance uses the last up to five visual
states and retains a GPT-5.5 reasoning target; three judge verdicts are kept in
results. These rows complete the SFT release to 100,000 judge instances.
This count is not a claim of 100,000 unique underlying trajectories because
some original-release trajectories have multiple screenshot or prompt views.