KoSum: 700 Recent Korean YouTube Videos with Multimodal Features

KoSum is a benchmark dataset containing 700 recent Korean YouTube videos uploaded between 2024 and 2026, spanning 14 fine-grained content categories. It provides per-second visual, audio, and text features aligned with YouTube Most-Replayed importance scores. The dataset was created by author iontail and last updated on June 14, 2026.

Use Cases

• Training multimodal video summarization models based on per-second visual, audio, and text features.
• Developing highlight detection algorithms based on YouTube Most-Replayed importance scores.
• Analyzing viewer-engagement-driven editing patterns across 14 content categories.
• Benchmarking models for Korean-language video understanding tasks.

Strengths

• Contains 700 recent videos, providing a substantial corpus for model training.
• Videos span 14 fine-grained content categories, offering diversity.
• Includes per-second multimodal features (visual, audio, text) aligned with engagement scores.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect temporal and platform bias inherent to YouTube videos from 2024-2026.

Provenance

Source

YouTube

Collection Method

Likely collected via YouTube API or web scraping.

Time Range

2024 to 2026

Freshness

Last updated 2026-06-14 12:53:38; freshness should be verified.

Geography

South Korea (based on language and content focus)