Dataset Card for OmniAssistBench

Dataset Summary

OmniAssistBench is a novel benchmark specially designed to evaluate Omni-Large Language Models (Omni-LLMs) under assistant-style, real-time video chat scenarios.

As multimodal models evolve from processing static images to understanding dynamic, real-time video streams, there is a pressing need for benchmarks that assess how these models perform as real-time visual assistants. OmniAssistBench bridges this gap by providing a comprehensive set of video-based interactions (approximately 67GB of high-quality video data), enabling researchers to evaluate the temporal understanding, contextual reasoning, and responsiveness of Omni-LLMs in chat environments.

Leaderboard can be found on the Project page.

Please refer to the GitHub page for detailed evaluation piplines and example codes.

Dataset Structure

Data Fields

Each instance in the dataset contains the following fields:

• video_full_path: The Hugging Face video object of the main video, corresponding to the video_name.
• video_name: The filename of the video, without the .mp4 extention.
• class: The task category of the video.
• source: The original source from which the video was collected.
• turns: Wether it is a single- or multi-turn question.
• question: The user's query directed at the assistant. (Note: The text provided in this field is only for human readability. During actual evaluation, user prompts have already been embedded into the video as audios, so there is no need to feed this text into the model as a separate input.)
• gt_answer: The Ground Truth answer of the question.
• key_points: Crucial points used for evaluating the model's response.
• example_video_full_path: The Hugging Face video object corresponding to the example_video.
• example_video: Prior knowledge provided to the model in the form of a reference video. (Note: This field is only applicable to the "Step Tracking" task.)
• example_text: Prior knowledge provided to the model in text format. (Note: The text provided in this field is only for human readability. During actual evaluation, this text has already been embedded into the video as subtitles, so there is no need to feed this text into the model as a separate input.)

Citation

If you use this dataset in your research or find it helpful, please consider citing our paper:

BibTeX:

@misc{sun2026omniassistbench,
      title={OmniAssistBench: Assistant-style Interaction Benchmark for Omni-LLMs}, 
      author={Xianyun Sun and Chaoyou Fu and Zhengye Zhang and Feiyang Duan and Qingyuan Cao and Yonghui Niu and Sihang Yuan and Ge Zhang and Caifeng Shan},
      year={2026},
      eprint={2608.21360},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.21360}, 
}