Datasets:

iontail
/

KoSum

license: cc-by-nc-4.0
task_categories:
  - summarization
language:
  - ko
tags:
  - video
  - summarization
  - highlight-detection
  - multimodal
  - most-replayed
  - youtube

Dataset Summary

KoSum is a Korean YouTube benchmark for video summarization and highlight detection. It contains 700 recent Korean YouTube videos uploaded between 2024 and 2026, spanning 14 fine-grained content categories. KoSum provides per-second visual, audio, and text features aligned with YouTube Most-Replayed importance scores.

The dataset is designed for multimodal video summarization, highlight detection, and analysis of viewer-engagement-driven editing patterns.

Project Page: iontail.github.io/kosum

Dataset Structure

The dataset contains metadata, multimodal features, ground-truth annotations, and standardized train/validation/test splits.

1. Metadata (`kosum_metadata.csv`)

Contains metadata for all 700 videos.

video_id: A short internal metadata ID.
youtube_id: The original YouTube video ID. This is the key used in all HDF5 files and kosum_split.json.
title: YouTube video title.
duration: Video duration in seconds.
views: YouTube view count.
category1: Broad content category.
category2: Fine-grained content category.
likes: YouTube like count.
comments: YouTube comment count.
category_id: YouTube category ID.
category: YouTube category name.
cap_lan: Caption language label.

KoSum has 14 fine-grained categories with 50 videos per category: action, animation, baseball, basketball, cooking, football, job, knowledge, lecture, review, romance, sketch, talk, and vlog.

2. Multimodal Features (`.h5` files)

Each feature file is an HDF5 file with 700 top-level datasets keyed by youtube_id. Each dataset has shape (T, 768), where T is the per-second sequence length for the video.

kosum_feat_visual_clip.h5: Visual features extracted with CLIP (openai/clip-vit-large-patch14).
kosum_feat_audio_ast.h5: Audio features extracted with Audio Spectrogram Transformer (MIT/ast-finetuned-audioset-10-10-0.4593).
kosum_feat_text_roberta.h5: Text features extracted from subtitles with XLM-RoBERTa (FacebookAI/xlm-roberta-base).

All feature datasets are stored as float32. The sequence lengths range from 383 to 1800 seconds.

Example HDF5 layout:

kosum_feat_visual_clip.h5
`-- {youtube_id}: float32[T, 768]

kosum_feat_audio_ast.h5
`-- {youtube_id}: float32[T, 768]

kosum_feat_text_roberta.h5
`-- {youtube_id}: float32[T, 768]

File Size & Downloading: Each feature file is approximately 2GB. Download time may vary depending on network conditions.

3. Ground Truth (`kosum_gt.h5`)

An HDF5 file containing ground-truth annotations for all 700 videos. Each youtube_id maps to an HDF5 group with the following keys:

change_points: Shot boundaries generated by KTS. Shape is (num_shots, 2), stored as int32.
gt_score: Per-second Most-Replayed importance scores. Shape is (T,), stored as float32.
gt_summary: Binary summary labels. Shape is (T,), stored as int8.

Each group also stores youtube_id, score_key, feature_key, category1, and category2 as HDF5 attributes.

Example HDF5 layout:

kosum_gt.h5
`-- {youtube_id}/
    |-- change_points: int32[num_shots, 2]
    |-- gt_score: float32[T]
    `-- gt_summary: int8[T]

4. Dataset Splits (`kosum_split.json`)

Contains standardized splits using youtube_id keys.

train_keys: 560 training videos.
val_keys: 70 validation videos.
test_keys: 70 test videos.

The split follows an 80/10/10 ratio and keeps each category2 balanced across train, validation, and test sets.

Loading Example

import h5py

video_id = "1_1aaQtxemQ"

with h5py.File("kosum_feat_visual_clip.h5", "r") as visual_h5:
    visual_feat = visual_h5[video_id][:]

with h5py.File("kosum_gt.h5", "r") as gt_h5:
    gt_score = gt_h5[video_id]["gt_score"][:]
    gt_summary = gt_h5[video_id]["gt_summary"][:]
    change_points = gt_h5[video_id]["change_points"][:]

Data Issues

If you find any data issue, such as missing keys, broken files, or metadata/annotation errors, please let us know through the Community.

Citation

If you use KoSum in your research, please cite:

@misc{lee2026kosum,
  title  = {KoSum: Beyond Detection: Most-Replayed Driven Multimodal Analysis of Korean YouTube Videos for Highlight Editing Guidance},
  author = {Lee, Chanhee and Jang, Jinho and Ha, Sungjun and Jung, Jinwoong},
  year   = {2026},
  note   = {2026 Spring DSC3028 Capstone Project Technical Report, Sungkyunkwan University}
}

Dataset Summary

Dataset Structure

1. Metadata (kosum_metadata.csv)

2. Multimodal Features (.h5 files)

3. Ground Truth (kosum_gt.h5)

4. Dataset Splits (kosum_split.json)