VideoMarathon: 9,700 Hours of Long Video Instruction-Following Data

VideoMarathon is a large-scale multimodal dataset containing approximately 9,700 hours of video paired with 3.3 million question-answer pairs. It was created by GarrickAI for academic research purposes and covers 22 diverse tasks across six fundamental topics. The dataset was last updated on 2026-05-25.

Use Cases

• Train video-language models for instruction-following based on the described QA pairs.
• Benchmark model performance on long video comprehension across the 22 task categories.
• Research temporal reasoning in videos based on the 'temporality' topic mentioned in the description.

Strengths

• Large scale with approximately 9,700 hours of total video duration.
• Contains 3.3 million instruction-following QA pairs.
• Covers a diverse range of 22 tasks across six fundamental topics.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

GarrickAI

Freshness

Last updated 2026-05-25 04:23:32; freshness should be verified.