TornadoLabs

Research arm of Velys Software. We release curated video datasets at production scale.

Mission

Make high-quality video datasets accessible to the AI research community. All datasets are processed via TornadoAPI's managed video extraction infrastructure.

Datasets

Videos are re-hosted on Hugging Face storage and served through direct HTTPS URLs — no more dead YouTube links, no more flaky downloader scripts.

Each dataset ships with:

• Raw MP4 videos at source quality, plus the source subtitles when available
• metadata.jsonl with one direct download URL per video
• The original academic annotations as captions.parquet
• A coverage note listing what could not be fetched (private / removed videos)

Figures as of 18 Sep 2026. Datasets are refreshed as more videos are delivered.

Still to come: HowTo100M is published as a 105,128-video subset of the 1.24M-video corpus and will grow in further batches. Kinetics-700 is next.

License

Videos remain the property of their uploaders; annotations belong to the original dataset authors (cite their papers). Re-hosted for reproducible research — see each dataset card for the source terms. For commercial use cases, contact TornadoAPI.

About

TornadoLabs is the research arm of Velys Software (Nice, France), a self-funded micro-enterprise. We operate TornadoAPI (managed video extraction infrastructure) and Tornado Data (custom datasets for AI labs).

→ Need custom video datasets?

Founded by Alexandre Sage.