Action100M

From the catalogue. Listed by one researcher from one source page.

• id: action100m
• kind: world model and video
• robot or device: internet instructional video (class: other)
• how collected: automated segmentation and captioning pipeline
• size as stated: ~100M segments from 1.2M videos (14.6 years); 10% preview on HF; 127,984 hours claimed, not released
• hours: not stated | hours only claimed: 127,984 | episodes: not stated
• year: 2026
• organisation: Meta FAIR
• licence: FAIR Noncommercial Research License (class: non commercial)
• access: open (class: open)
• link (dataset): https://huggingface.co/datasets/facebook/action100m-preview
• Hugging Face repository: facebook/action100m-preview, 1,122 downloads in 30 days, https://huggingface.co/datasets/facebook/action100m-preview
• other names: Action100M preview
• part of: HowTo100M. Its hours are counted there, not added again to the index's total
• note: Annotations only; segmentation uses V-JEPA 2 embeddings.

Confirm the licence at the link before relying on it.