Datasets:

BomiaoWang
/

DVBench

license: cc-by-nc-sa-4.0
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - video
  - chart
  - multimodal
  - benchmark
pretty_name: DVBench
configs:
  - config_name: default
    data_files:
      - split: test
        path: test.parquet

DVBench

DVBench is a benchmark for evaluating multimodal large language models on data videos, a storytelling medium that combines dynamic charts with structured narratives.

300 data videos
1,000 human-verified question-answer pairs
Five dimensions: Narrative, Animation, Chart Perception, Chart Reasoning, and Alignment

Dataset Structure

Each row contains:

question_id: unique question identifier
question_type: EM, MCQ_single, MCQ_multiple, or Open_ended
video: video identifier matching videos/<video>.mp4
dimension: evaluation dimension
question: question or subtitle-cloze context
answer: reference answer
distractor1, distractor2, distractor3: distractor options for single-choice and multiple-choice questions
chart_type: chart category for visual questions in Animation, Chart Perception, and Chart Reasoning
animation_editorial_layer: editorial-layer category for Animation questions
chart_reas_type: data-insight category for Chart Reasoning questions
alignment_semantic_label: semantic category (Data Insight or Data Context) for Alignment questions

Loading

from datasets import load_dataset

dataset = load_dataset("BomiaoWang/DVBench", split="test")
print(dataset[0])

Dataset Statistics

(a) Question distribution. Distribution of questions across the five evaluation dimensions. Chart Reasoning receives the largest share because viewers primarily focus on underlying data insights rather than superficial graphical elements, and current models remain more vulnerable in reasoning than in basic perception.
(b) Chart types. Chart-type distribution of visual questions and its dimension-specific breakdown. Visual questions include Animation, Chart Perception, and Chart Reasoning. Bar and line charts are the most prevalent, while the remaining questions cover a diverse long tail of chart types.
(c) Video length and topics. Video lengths range from under 30 seconds to approximately 37 minutes and are grouped as Short (≤3 minutes), Medium (3–6 minutes), and Long (>6 minutes). The videos cover ten real-world topics, with Politics & Society and Economy & Finance as the two largest categories.

Videos and Rights

The videos are third-party works and are not relicensed under the annotation license. Copyright and related rights remain with their creators and rightsholders. Access and use must comply with applicable law, source-platform terms, and any restrictions imposed by the rightsholder.

License

The benchmark annotations and metadata are released under CC BY-NC-SA 4.0. They are intended for non-commercial academic research and evaluation.

Citation

If you find DVBench useful, please cite:

@inproceedings{wang2026dvbench,
  title={DVBench: Benchmarking MLLMs for Understanding Dynamic Charts and Narratives in Data Videos},
  author={Wang, Bomiao and Shao, Zekai and Lan, Jiexiang and Fu, Xiaoliang and Zeng, Xingchen and Chen, Siming},
  booktitle={Findings of the Association for Computational Linguistics: EMNLP 2026},
  year={2026}
}

Repository

Code and evaluation utilities: https://github.com/BomiaoWang/DVBench