Hugging Face's logo

Datasets:
198808xc
/
drama_bench_data

Tasks:
Video-Text-to-Text
ArXiv:
License:
Dataset card Files
xet
 Community
1
Dataset Viewer
The dataset viewer is not available for this subset.
Job manager crashed while running this job (missing heartbeats).

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

DramaVQA

DramaVQA is a large-scale benchmark for evaluating detailed understanding of ultra-long TV dramas. It contains approximately 20,000 question-answering pairs for evaluation, plus a separate training set with 86,736 questions, spanning 15 dramas totaling about 606 hours.

Data Structure

  • Single-point Full: 20,000 questions (12 dramas)
  • Single-point Mini: 2,000 questions (subset of Full)
  • Multi-point: 3,200 questions
  • Multi-hop: 1,200 questions
  • Training set: 86,736 questions (3 separate dramas)

Question categories cover 7 major types and 28 subcategories: visual perception, spatial understanding, action and interaction, event understanding, temporal understanding, relation & logical reasoning, and counting.

Sample Usage

To download the media for one evaluation drama (e.g., The Knockout), use:

git clone https://github.com/xinyuehuo/DramaVQA.git
cd DramaVQA
python -m pip install --upgrade huggingface_hub
hf download 198808xc/drama_bench_data --repo-type dataset \
  --include "the_knockout/*" --local-dir data/drama_bench_data

Change --include for another drama, or omit it to download all media.

License

Released under the MIT License.
Downloads last month
5,706
Total file size:
2.24 TB

Paper for 198808xc/drama_bench_data