Cambrian-S-3M: 3 Million Video Instruction Tuning Records

Cambrian-S-3M is a collection of approximately 3 million video instruction tuning records developed by nyu-visionx for the third training stage of the Cambrian-S multimodal model. Released in early 2026, the dataset aggregates video-text pairs from Cambrian-S-3M, LLaVA-Video-178K, and LLaVA-Hound (ShareGPTVideo).

Use Cases

• Instruction tuning of multimodal models using video-text pairs
• Benchmarking video understanding capabilities against LLaVA-Video-178K subsets
• Training conversational agents to describe temporal events in ShareGPTVideo sequences

Strengths

• Aggregates 3,000,000 video instruction records
• Combines established benchmarks including LLaVA-Video-178K and LLaVA-Hound
• Released under Apache 2.0 license

Limitations

• Requires approximately 5 TB of disk space for storage
• Potential for label noise or format inconsistencies across the three source datasets

Provenance

Source

nyu-visionx (Arxiv 2511.04670)

Collection Method

Aggregation of existing open-source video instruction datasets

Freshness

Last updated January 2026.