Videomind Dataset

VideoMind-SFT contains 481,000 video-annotation pairs and a 210,000-record Grounder subset released by yeliudev in early 2026. The collection provides videos in both original formats and compressed versions at 3 FPS and 480p resolution without audio for efficient model training.

Use Cases

• Supervised fine-tuning of video-language models using the 481K SFT records
• Temporal grounding evaluation using the 210K Grounder subset
• Benchmarking video highlight detection using the QVHighlights directory

Strengths

• Contains 481,000 SFT records for large-scale training
• Includes 210,000 grounding-specific annotations
• Provides dual video formats including 3 FPS/480p compressed versions
• Linked to Arxiv publication 2503.13444

Limitations

• Compressed video versions lack audio tracks entirely
• Low frame rate (3 FPS) in compressed files may limit motion-sensitive analysis
• Column-level documentation is not provided in the primary repository description

Provenance

Source

yeliudev (Arxiv 2503.13444)

Freshness

Last updated January 2026.

Geography

United States