You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
The LongVQU-Eval dataset contains identifiers to videos originally sourced from MLVU, LongVideoBench, and LongVideo-Reason-Eval for data collection purposes. LongVQU-Eval does not own the content linked within this dataset; all rights and copyright belong to the respective dataset owners and original channel creators. Ensuring compliance with platform terms and conditions is the responsibility of these source datasets. By accessing this dataset, you acknowledge and agree to the following terms:
Log in or Sign Up to review the conditions and access this dataset content.
LongVQU-Eval
Test set of the LongVQU-Eval Challenge on long-term video quality understanding
🏆 Challenge (Codabench) | 📊 LongVQUBench | 🌐 Project Page | 💻 GitHub
Dataset Description
LongVQU-Eval contains the test videos of the LongVQU-Eval Challenge. The challenge evaluates long-term video quality understanding in large vision-language models (LVLMs) and follows the task hierarchy of LongVQUBench. Questions, the submission format and the leaderboard are provided on the Codabench challenge page.
The videos are derived from MLVU, LongVideoBench and LongVideo-Reason-Eval, with spatial and temporal quality distortions applied to selected segments.
Task Hierarchy
| Level | Name | Sub-tasks |
|---|---|---|
| LQU | Local Event Quality Understanding | Classification, Detection, Localization, Severity and comfort |
| CQR | Cross-Event Quality Reasoning | Comparison, Cumulative effect, Integration, Temporal relationship |
| GQU | Global Quality Understanding | Dominant factors, Overall quality rating, Stability and consistency, Temporal trend |
Repository Structure
The videos are packed into 1 GB tar parts, grouped by video group, level and sub-task:
GROUP_1/
├── CQR/
│ ├── CQR_Comparison/
│ │ ├── CQR_CQR_Comparison_001.tar
│ │ ├── CQR_CQR_Comparison_002.tar
│ │ └── ...
│ ├── CQR_Cumulative_effect/
│ ├── CQR_Integration/
│ └── CQR_Temporal_relationship/
├── GQU/
│ ├── GQU_Dominant_factors/
│ ├── GQU_Overall_quality_rating/
│ ├── GQU_Stability_and_consistency/
│ └── GQU_Temporal_trend/
└── LQU/
├── LQU_Classification/
├── LQU_Detection/
├── LQU_Localization/
└── LQU_Severity_and_comfort/
| Group | Videos | Tar parts | Size |
|---|---|---|---|
| GROUP_1 | 300 | 252 | ~247 GB |
Important: the tar parts of one sub-task folder are consecutive pieces of a single tar stream. A single part cannot be extracted on its own; all parts of a folder must be concatenated in numeric order before extraction (see below).
Download
Request access on this page first, then log in with a token from an account that has been granted access:
pip install -U "huggingface_hub[hf_xet]"
hf auth login
Download the whole group:
hf download Aarna004/LongVQU-Eval --repo-type dataset \
--include "GROUP_1/**" --local-dir ./LongVQU-Eval
Or download a single sub-task:
hf download Aarna004/LongVQU-Eval --repo-type dataset \
--include "GROUP_1/CQR/CQR_Comparison/*" --local-dir ./LongVQU-Eval
Extracting the Videos
Concatenate the parts of each sub-task in numeric order and extract the stream:
cd LongVQU-Eval
mkdir -p ../videos/GROUP_1
for dir in GROUP_1/*/*/; do
echo "Extracting ${dir}"
cat $(ls "${dir}"*.tar | sort -V) | tar -xf - -C ../videos/GROUP_1
done
The videos are restored as videos/GROUP_1/<LEVEL>/<SUB-TASK>/<video_idx>.mp4,
e.g. videos/GROUP_1/CQR/CQR_Comparison/1070.mp4.
License
LongVQU-Eval is released under the CC BY-NC-SA 4.0 license, for non-commercial research use only. All rights and copyright of the original videos belong to the respective source datasets (MLVU, LongVideoBench, LongVideo-Reason-Eval) and the original content creators.
Citation
@inproceedings{nema2026longvqubench,
title = {LongVQUBench: Benchmarking Long-Term Video Quality Understanding of Vision-Language Models},
author = {Nema, Arpita and Zhu, Hanwei and Zhang, Xi and Lin, Weisi},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026},
url = {https://longvqubench.github.io}
}
Contact
For questions about the dataset or the challenge, please open a discussion on this repository or use the forum of the Codabench challenge.
- Downloads last month
- 511