OneJev-Data

<p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="assets/banner-dark.svg"> <img alt="OneJev-Data, the training data of OneJev" src="assets/banner.svg" width="100%"> </picture> </p> <p align="center"> <a href="https://huggingface.co/collections/OmniJev/onejev"><img alt="Hugging Face" src="https://img.shields.io/badge/Hugging_Face-OneJev-FFD21E?style=flat-square&logo=huggingface&logoColor=black"></a> <a href="https://huggingface.co/spaces/BradNLP/OneJev"><img alt="Demo" src="https://img.shields.io/badge/Demo-Try_it-F83800?style=flat-square&logo=gradio&logoColor=white"></a> <a href="https://github.com/OmniJev/OneJev"><img alt="GitHub" src="https://img.shields.io/badge/GitHub-OmniJev%2FOneJev-181717?style=flat-square&logo=github&logoColor=white"></a> <a href="https://omnijev.github.io/OneJev/"><img alt="Website" src="https://img.shields.io/badge/Website-OneJev-0A84FF?style=flat-square&logo=googlechrome&logoColor=white"></a> <a href="https://github.com/OmniJev/awesome-jev-gallery"><img alt="Awesome JEV" src="https://img.shields.io/badge/Awesome-JEV-FC60A8?style=flat-square&logo=awesomelists&logoColor=white"></a> </p> The training data of [OneJev](https://huggingface.co/collections/OmniJev/onejev): 94,707 typed questions about screens, photos, videos and text, each with its answer. This is 95.5% of the rows OneJev was trained on; rows whose sources do not allow redistribution are left out. **Try the demo data:** [Download 100 examples with images (23 MB)](https://huggingface.co/datasets/OmniJev/OneJev-Data/resolve/main/sample/sample-100.parquet?download=true). <p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="assets/groups-dark.svg"> <img alt="Rows per group: GUI agent runs 33,405, text 16,849, images 13,559, short videos 12,115, rules 12,000, long videos 6,779" src="assets/groups.svg" width="100%"> </picture> </p> ## Try 100 examples **OneJev-Data · 100-example starter sample** is a small dataset for exploring what OneJev learns from and trying your own data-loading code. It covers GUI actions, images, short and long videos, text, and rules. Each example contains a situation (`state`), a question (`question`), and its expected answer (`target`). | What you want to do | Where to start | |---|---| | Read the 100 examples without writing code | [Text preview (JSON)](sample/sample-100-preview.json) | | Load the examples with their original images and video frames | [100-example sample (Parquet, 23 MB)](sample/sample-100.parquet), or the code below | | Try the model interactively | [OneJev online demo](https://huggingface.co/spaces/BradNLP/OneJev) | **Parquet is the file format**: it stores the records and their images together in one file. The sample has the same schema as the full dataset. The JSON preview is readable text with expanded fields; it contains media references, but the actual images and video frames are included only in the Parquet download. To load just the 100-example sample: ```python from datasets import load_dataset ds = load_dataset("OmniJev/OneJev-Data", "sample", split="train") row = ds[0] ``` These 100 rows are a uniform random sample of the released training rows (seed 0), intended for inspecting the format and testing your pipeline. They are part of the training set, not a held-out evaluation set, and retain their source licenses. ## Use the full dataset The full dataset contains 94,707 rows in 18 Parquet files, totaling **17.7 GB** (about 984 MB per file). The files include image bytes and sampled video frames, so they can be used without downloading media separately. To iterate over the full dataset without downloading it all first, use streaming: ```python from datasets import load_dataset ds = load_dataset("OmniJev/OneJev-Data", "default", split="train", streaming=True) row = next(iter(ds)) ``` Each row has a `state` with `<image:N>` and `<video:N>` placeholders, a `question`, its `target`, and the pictures in `images`. To train OneJev on it, see [GitHub](https://github.com/OmniJev/OneJev#training). ## Sources 127 public datasets. The largest in each group: | Group | Sources | |---|---| | GUI agent runs | [AMEX](https://huggingface.co/datasets/Yuxiang007/AMEX), [Multimodal-Mind2Web](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web), [OSWorld-Verified](https://huggingface.co/datasets/xlangai/ubuntu_osworld_verified_trajs), [AgentRewardBench](https://huggingface.co/datasets/McGill-NLP/agent-reward-bench), [MisActBench](https://huggingface.co/datasets/osunlp/MisActBench) and 4 more | | Images | [GQA](https://cs.stanford.edu/people/dorarad/gqa/download.html), [ImageRewardDB](https://huggingface.co/datasets/zai-org/ImageRewardDB), [GenAI-Bench](https://huggingface.co/datasets/TIGER-Lab/GenAI-Bench), [FGVC-Aircraft](https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/), [CLEVR](https://cs.stanford.edu/people/jcjohns/clevr/) and 30 more | | Short videos | [VideoFeedback](https://huggingface.co/datasets/TIGER-Lab/VideoFeedback), [STAR](https://prior.allenai.org/projects/charades), [MSR-VTT](https://www.microsoft.com/en-us/research/publication/msr-vtt-a-large-video-description-dataset-for-bridging-video-and-language/), [Diving48](http://www.svcl.ucsd.edu/projects/resound/dataset.html), [CLEVRER](http://clevrer.csail.mit.edu/) and 3 more | | Long procedural videos | [Breakfast](https://serre.lab.brown.edu/breakfast-actions-dataset.html), [CaptainCook4D](https://captaincook4d.github.io/captain-cook/) | | Text | [SWE-agent trajectories](https://huggingface.co/datasets/nebius/SWE-agent-trajectories), [DTap-Bench](https://huggingface.co/datasets/AI-Secure/DTap-Bench-Agent-Trajectories), [SWE-bench experiments](https://github.com/swe-bench/experiments), [Terminal-Bench 2](https://huggingface.co/datasets/harborframework/terminal-bench-2-leaderboard) and 66 more | | Rules | [typed-decisions-synth](https://huggingface.co/datasets/n4ze3m/typed-decisions-synth), [jevlite](https://huggingface.co/datasets/vagmi/jevlite_dataset) and our own set written with Qwen3.5-27B | All 127 with links and licenses are in [licenses.csv](licenses.csv). ## Citation Please cite OneJev-Data and OneJev, and the sources whose rows you use. ```bibtex @misc{onejevdata2026, title = {{OneJev-Data}: The Training Data of {OneJev}}, author = {{OmniJev Team}}, year = {2026}, howpublished = {\url{https://huggingface.co/datasets/OmniJev/OneJev-Data}} } @misc{onejev2026, title = {{OneJev}: A Multimodal System One Decision Model}, author = {{OmniJev Team}}, year = {2026}, howpublished = {\url{https://github.com/OmniJev/OneJev}} } ``` ## License Each row keeps the license of its source, given in the `license` column and in [licenses.csv](licenses.csv). Some sources allow research or non-commercial use only. To have a source removed, open an issue on [GitHub](https://github.com/OmniJev/OneJev/issues).