VideoMix22M (V-JEPA 2 pre-training set)

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

• id: videomix22m-v-jepa-2-pre-training-set
• kind: video for world models
• robot or device: none (internet video) (class: other)
• how collected: curated mix of public internet video datasets
• size as stated: 22 million samples including 1 million images; 1,000,000 hours claimed, not released
• hours: not stated | hours only claimed: 1,000,000 | episodes: not stated
• year: 2025
• organisation: Meta FAIR
• licence: not stated (class: not stated)
• access: not released as a set; component datasets are public (class: not stated)
• link (paper): https://arxiv.org/abs/2506.09985
• note: Mostly general YouTube video rather than egocentric capture; the paper's table lists 1.6M source hours for YT-Temporal-1B and 134K for HowTo100M before curation.

Confirm the licence at the link before relying on it.