sonalkum/MMOU-Videos

MMOU Videos

This repository solely hosts the raw video files and captions for the MMOU benchmark. For the actual dataset, questions, evaluation scripts, and full documentation, please visit nvidia/MMOU.

What This Repository Contains

This repository is the storage companion for MMOU. It contains the heavy media assets needed to run the benchmark:

• Videos: 9,049 root-level real-world videos stored as MP4 files and sampled at 720p.
• Captions (MMOU_Captions.jsonl): detailed audio-visual captions describing visuals, speech, sound, and music.

The caption file is intended only for the open-ended evaluation judge. It should not be provided to models being evaluated on MMOU.

What This Repository Does Not Contain

This repository does not contain the benchmark logic itself. It does not host:

• question-answer annotations
• multiple-choice options
• evaluation scripts
• leaderboard tables
• methodological documentation

For all of that, go to nvidia/MMOU.

File Layout

Videos are stored as MP4 files named by source video ID:

<youtube_id>.mp4

Example:

Rja2JeIaH_c.mp4

Test Mini Videos

The test/ directory contains the MP4 videos used by the MMOU Test Mini split released in nvidia/MMOU. File names follow the same convention as the main video set: each file is named by its 11-character YouTube video ID, for example VIDEO_ID.mp4.

This directory currently contains 2,832 MP4 files corresponding to the available videos for the 5,000-question Test Mini split.

Usage

You can load the video repository with the Hugging Face datasets library:

from datasets import load_dataset

videos = load_dataset("sonalkum/MMOU-Videos", split="train", streaming=True)
sample = next(iter(videos))

print(sample)

To stream the Test Mini videos, load the test split:

from datasets import load_dataset

test_videos = load_dataset("sonalkum/MMOU-Videos", split="test", streaming=True)
sample = next(iter(test_videos))

print(sample)

If you need the caption file locally, load it separately as JSONL:

from datasets import load_dataset

captions = load_dataset("json", data_files="MMOU_Captions.jsonl", split="train")
print(captions[0])

Notes for Evaluation

• Use the videos in this repository together with the benchmark annotations from nvidia/MMOU.
• Use MMOU_Captions.jsonl only for the open-ended judge pipeline.
• Do not feed the captions to the model under evaluation if you want a valid MMOU result.

License

Apache License 2.0

Citation

If you use MMOU, please cite:

@misc{goel2026mmoumassivemultitaskomni,
      title={MMOU: A Massive Multi-Task Omni Understanding and Reasoning Benchmark for Long and Complex Real-World Videos}, 
      author={Arushi Goel and Sreyan Ghosh and Vatsal Agarwal and Nishit Anand and Kaousheik Jayakumar and Lasha Koroshinadze and Yao Xu and Katie Lyons and James Case and Karan Sapra and Kevin J. Shih and Siddharth Gururani and Abhinav Shrivastava and Ramani Duraiswami and Dinesh Manocha and Andrew Tao and Bryan Catanzaro and Mohammad Shoeybi and Wei Ping},
      year={2026},
      eprint={2603.14145},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.14145}, 
}