| --- |
| license: apache-2.0 |
| pretty_name: OS-Omni Benchmark |
| task_categories: |
| - text-generation |
| tags: |
| - os-omni |
| - benchmark |
| - desktop-agents |
| - gui-agents |
| - android |
| - ios |
| - linux |
| - macos |
| - windows |
| - cross-vm |
| - croissant |
| size_categories: |
| - 100<n<1K |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/tasks.parquet |
| --- |
| |
| # OS-Omni Benchmark |
|
|
| OS-Omni is a cross-platform benchmark for evaluating agents that operate graphical operating-system environments. This dataset repository contains the static benchmark task definitions and supporting assets used to configure and evaluate OS-Omni tasks. |
|
|
| ## Contents |
|
|
| - `data/tasks.parquet`: tabular task index for Hugging Face Dataset Viewer and Croissant generation. |
| - `data/tasks.jsonl`: JSON Lines copy of the same task index. |
| - `metadata/tasks.parquet`: duplicate task index kept with the metadata bundle. |
| - `metadata/tasks.jsonl`: duplicate JSON Lines task index kept with the metadata bundle. |
| - `metadata/summary.json`: aggregate counts by platform, oracle status, and difficulty. |
| - `croissant/os_omni_benchmark_croissant_rai.json`: hand-authored Croissant metadata with NeurIPS-style minimal RAI fields. |
| - `tasks/`: original YAML task specifications, including initialization and evaluation parameters. |
| - `assets/`: static assets needed by benchmark setup and documentation. |
| - `files/`: small static files used by desktop tasks. |
| - `apps/quickshop/shared/seed-data.json`: seed catalog data for QuickShop tasks. |
| - `executions/suites/`: benchmark suite selection files. |
| - `MANIFEST.sha256`: SHA256 checksums for uploaded files. |
|
|
| The VM and emulator images are intentionally not included in this repository. They should be hosted as separate large artifacts and referenced from the OS-Omni code repository or paper. |
|
|
| ## Task Index Schema |
|
|
| Each row in `metadata/tasks.parquet` describes one YAML task file. |
|
|
| - `task_id`: stable task identifier from the YAML file. |
| - `platform`: target platform, such as `android`, `ios`, `linux`, `macos`, `windows`, or `cross_vm`. |
| - `is_oracle`: whether the file is an oracle task variant. |
| - `difficulty`: task difficulty when specified or inferable from the filename. |
| - `app_name`: primary application or domain inferred from metadata or task id. |
| - `instruction`: user-facing task instruction. |
| - `task_yaml_path`: path to the original YAML task definition in this repository. |
| - `initializer_id`: Python initializer entry point. |
| - `evaluator_id`: Python evaluator entry point. |
| - `tags`: JSON-encoded list of task tags. |
| - `metadata_json`, `init_json`, `evaluation_json`: JSON-encoded structured fields from the YAML file. |
| - `sha256`: checksum of the original YAML task file. |
|
|
| ## Use |
|
|
| Clone or download this dataset alongside the OS-Omni code repository. The task YAML files are intended to be consumed by the OS-Omni runner, while `metadata/tasks.parquet` is intended for inspection, filtering, documentation, and metadata tooling. |
|
|
| ## Responsible AI Notes |
|
|
| This benchmark is intended for measuring GUI-agent behavior in controlled benchmark environments. It is not intended to evaluate real user accounts, production systems, or sensitive personal data. Some tasks contain synthetic names, emails, phone-like strings, credentials for local benchmark applications, and operational scenarios designed as test fixtures. These should not be interpreted as real personal records. |
|
|
| Known limitations include platform and application coverage gaps, dependence on VM/emulator setup quality, possible evaluator brittleness, and distributional differences from real-world user workflows. Model performance on OS-Omni should not be treated as a complete measure of general operating-system autonomy or safety. |
|
|
| ## License |
|
|
| Apache-2.0. See `LICENSE`. |
|
|
