Access requests are reviewed manually, usually within one week. Please provide your affiliation and intended non-commercial research or educational use.

SportsTime

SportsTime is a long-form sports video question answering benchmark for temporal compositional reasoning, accepted to ECCV 2026.

It contains 14,326 open-ended QA pairs with 50,000+ step-wise temporal evidence annotations across 1,575 videos and five team sports: basketball, American football, ice hockey, soccer, and volleyball.

SportsTime overview

Dataset

This Hugging Face dataset repository provides the annotation files, official train/test split, and video files. Download access is gated and reviewed manually, usually within one week.

Configuration	Train	Test
Chinese (zh)	11,540	2,786
English (en)	11,540	2,786

Video files are stored under videos/{sport}/{full|highlight}/{video_id}.mp4. See metadata/video_manifest.json for the complete file list and sizes.

Usage

from datasets import load_dataset

zh = load_dataset("Ustiniansy/SportsTime", "zh")
en = load_dataset("Ustiniansy/SportsTime", "en")

Each sample contains the question, answer, Chain-of-Time reasoning, temporal references, sport, video type, language, and the associated video_id.

Citation

@article{cao2026sportstime,
  title   = {Towards Temporal Compositional Reasoning in Long-Form Sports Videos},
  author  = {Cao, Siyu and Zhang, Lu and Zeng, Ruizhe and Liu, Zhi-yong},
  journal = {arXiv preprint arXiv:2604.22226},
  year    = {2026},
  eprint  = {2604.22226},
  archivePrefix = {arXiv},
  url     = {http://arxiv.org/abs/2604.22226}
}

License

SportsTime is released under the Creative Commons Attribution-NonCommercial 4.0 International license.

Downloads last month: 208

Total file size:

657 GB

Paper for Ustiniansy/SportsTime

Towards Temporal Compositional Reasoning in Long-Form Sports Videos

Paper • 2604.22226 • Published Apr 24

Request Access to SportsTime