Hugging Face's logo

Datasets:
davidchan
/
svld

Tasks:
Image-Text-to-Text
Image-to-Text
Video-Text-to-Text
Modalities:
Image
Video
Text
Languages:
English
Size:
1M<n<10M
ArXiv:
Tags:
multimodal
vision-language
social-media
image
video
text
Libraries:
Datasets
License:
Dataset card Files
xet
Community

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.

Social Vision and Language Dataset (SVLD)

Original Paper: A Dataset and Benchmarks for Multimedia Social Analysis (2020)
Authors: Bofan Xue, David Chan, John Canny
Institution: University of California, Berkeley

📌 Overview

The Social Vision and Language Dataset (SVLD) is a large-scale multimodal social media dataset designed to support research in:

  • Vision–language modeling
  • Multimodal fusion
  • Social signal prediction
  • Comment-tree modeling
  • Temporal social dynamics
  • Content popularity prediction

SVLD combines images, videos, text, social engagement signals, and full comment trees within the same context, enabling joint modeling across modalities in realistic, in-the-wild social media settings.

📦 Current Release (S3 Shard Edition)

The dataset is currently distributed as:

1961 daily shards

Each shard corresponds approximately to one day of collected data.

⚠ Important Notice

Due to long-term storage issues and partial data corruption:

  • This release may not contain the full original dataset
  • Some days, posts, media files, or metadata may be missing
  • Total dataset size may vary

Researchers are strongly encouraged to:

  • Recompute dataset statistics locally
  • Avoid assuming counts from the original publication
  • Design pipelines that tolerate partial or missing data

🧩 Dataset Structure

Each Post May Contain

  • One or more images
  • One or more videos
  • Optional per-media descriptions
  • A natural language title
  • User-provided tags
  • Social signals (upvotes, downvotes, favorites, views)
  • Timestamp
  • A full comment forest

Each Comment May Contain

  • Text
  • Images
  • GIFs or videos
  • Recursive replies (tree structure)

🎯 Modalities

SVLD supports research across:

  • Images (posts + comments)
  • Videos (posts + comments)
  • Text (titles, descriptions, comments)
  • Social Metrics (votes, favorites, views)
  • Tags (user-generated)
  • Tree Structure (comment forests)
  • Temporal Data (timestamps)

🔬 Research Directions

SVLD enables work in:

  • Multimodal fusion architectures
  • Image/video + language modeling
  • Popularity and engagement prediction
  • Social dynamics modeling
  • Tag and metadata prediction
  • Comment tree reasoning
  • Temporal distribution analysis
  • Multimodal retrieval
  • Content moderation research

⚙ Data Quality Notes

  • Some media files may be unavailable
  • Some shards may be incomplete
  • Social metrics reflect snapshot-at-scrape time
  • Engagement distributions are heavily long-tailed
  • Content reflects real-world social media (unfiltered, in-the-wild)

📖 Citation

If you use SVLD, please cite:

Xue, B., Chan, D., & Canny, J. (2020).
A Dataset and Benchmarks for Multimedia Social Analysis.
arXiv:2006.08335

📜 License & Usage

This dataset is intended for academic research use only.
Users are responsible for complying with platform terms and ethical research standards.

Downloads last month
-

Paper for davidchan/svld