NVIDIA Cosmos training video corpus

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

• id: nvidia-cosmos-training-video-corpus
• kind: video for world models
• robot or device: none (mixed web and proprietary video) (class: other)
• how collected: proprietary and public video curated into clips
• size as stated: about 100M clips for pre-training and about 10M for fine-tuning; 20,000,000 hours claimed, not released
• hours: not stated | hours only claimed: 20,000,000 | episodes: not stated
• year: 2025
• organisation: NVIDIA
• licence: not released (class: unclear)
• access: not released (class: not stated)
• link (paper): https://arxiv.org/abs/2501.03575
• note: Internal training corpus for the Cosmos world foundation models; the models are public but the video is not.

Confirm the licence at the link before relying on it.