Hugging Face's logo

Datasets:
Lanxingxuan
/
Video_Crop_Data

Tasks:
Visual Question Answering
Video Classification
Modalities:
Image
Video
Formats:
imagefolder
Languages:
English
Size:
10K - 100K
Tags:
video
tool-use
agentic
sft
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Video Crop Data

Agentic video temporal-crop trajectories derived from LVBench multiple-choice QA.

Contents

  • annotations/accepted_trajectories.jsonl: 1,000 complete reviewed trajectories.
  • annotations/video_crop_sft.jsonl: 2,591 prefix-recovery SFT rows.
  • media/videos/: 1 FPS, 180p short-side capped, silent overview videos.
  • media/frames/: 360p short-side capped timestamped crop observations; source media is never upscaled.
  • metadata/media_manifest.jsonl: repository path, byte size, and SHA-256 for every media file.
  • metadata/production_summary.json: production and review statistics.

Paths in both annotation files are relative to the repository root. The complete overview is the initial video input. A temporal crop tool returns timestamped JPEG images as the next-round observation.

The source benchmark media and questions retain their original upstream terms. The repository license metadata does not override upstream dataset or media licensing requirements.
Downloads last month
631
Total file size:
4.5 GB