CedPei/TempCloze

TempCloze

Paper: TempCloze: Can Video-LLMs Identify the Missing Middle?

TempCloze is a video cloze benchmark for evaluating whether Video-LLMs can identify the missing middle of a video from its beginning and ending context.

This repository provides metadata for 1,521 videos from seven sources. Each row identifies one source video and the temporal boundaries of its missing segment. The benchmark code and evaluation instructions are available in the official GitHub repository.

Dataset structure

The test split contains the following fields:

The three video segments are determined directly from the source video and timestamps:

• —beginning: [0, gap_start)
• —ground truth: [gap_start, gap_end)
• —ending: [gap_end, video_end]

Example:

{"source":"care","id":"v_00000203_5.mp4","gap_start":5.6,"gap_end":10.4}

Dataset composition

Loading the metadata

from datasets import load_dataset

dataset = load_dataset("CedPei/TempCloze", split="test")
print(dataset[0])

Citation

@article{pei2026tempcloze,
  title   = {TempCloze: Can Video-LLMs Identify the Missing Middle?},
  author  = {Pei, Wenqi and Zhao, Henry Hengyuan and Liu, Yilai and Meng, Jiahao and Chen, Han and Wang, Ziyu and Du, Hongyang},
  journal = {arXiv preprint arXiv:2609.01515},
  year    = {2026}
}