mxmprhd/youtube-shorts-dataset

YouTube Shorts Cover-Video Corpus

525 YouTube Shorts, 18 categories, 45 creators. Cover-video corpus for StegoReels, a robust neural steganography project for social video.

Metadata only — id, URL, creator, category, split, duration, resolution, engagement stats — not the video files. Why: Considerations. How to get the actual clips: Reproducing.

Summary

3–60s, vertical, H.264/AAC. Categories: challenge/stunts, sports trick shots, high-speed science, gaming, tech review, cooking, DIY/crafts, VFX, travel, animation/edu, ASMR, dance, comedy, fitness, music, animals, nature/wildlife, beauty, education, kids'.

Duration 5–60s, mean 35.3s.

Dataset Structure

Data Fields

field	type	description
id	string	YouTube video id
title	string	Video title
webpage_url	string	Canonical youtube.com/watch?v=... URL
source_query	string	Channel /shorts URL this clip was discovered through
uploader	string	Channel display name
channel_id	string	YouTube channel id
upload_date	string	YYYYMMDD
duration_s	int	Duration in seconds
width, height	int	Final encoded resolution
fps	float	Frame rate
view_count, like_count	int	Snapshot at scrape time
license	string or null	YouTube-reported license, if any
tags	list[string]	Uploader-supplied video tags
ext	string	Always mp4
category	string	One of 18 curated content categories

split is not a column — it's encoded by which file a row lives in (data/train-*.parquet, data/validation-*.parquet, data/test-*.parquet), per the standard Hugging Face convention.

Data Splits

Split by channel, per category: 1 channel → train; 2 → train+val; 3+ → train, val, test (one each, rest train). Every category keeps at least one creator out of training, so val/test measure generalization to unseen creators, not memorized style. Train channels get the full clip quota, val/test get ~35% of it.

category	train	validation	test	total
animals_pets	20	7	0	27
animation_edu	20	7	0	27
asmr	20	7	7	34
beauty_fashion	20	2	0	22
challenge_stunts	20	7	0	27
comedy_sketch	20	7	7	34
cooking_food	20	7	7	34
dance	20	7	0	27
diy_crafts	20	7	7	34
fitness	20	7	7	34
gaming	5	7	0	12
high_speed_science	20	7	7	34
music_performance	19	7	0	26
nature_wildlife	20	7	0	27
sports_trick_shots	20	7	7	34
tech_review	20	7	7	34
travel_adventure	20	7	4	31
vfx_magic	20	7	0	27

Dataset Creation

Creators resolved via YouTube search and verified to have a working /shorts tab (hand-typed @handles are unreliable — renamed or missing often enough). Clips pulled with yt-dlp, filtered to 3–60s + vertical, forced to H.264/AAC (YouTube's highest-res Shorts are otherwise AV1, which QuickTime and other players can't decode on many Macs).

category is assigned per channel (the creator's primary content type), not per video.

Considerations for Using the Data

• No video files. Clips belong to their original creators; this indexes public metadata only. Fetch the actual video with scripts/ here (subject to YouTube's Terms of Service).
• Selection bias: channels picked by hand for diversity, not randomly sampled — diverse, not representative.
• Metadata drift: view/like counts are a scrape-time snapshot; channels can rename, delete, or drop their Shorts tab afterward.

Reproducing / downloading the videos

pip install yt-dlp pandas pyarrow
python scripts/scrape_dataset.py --channels-file config/channels.csv --target-total 525

Re-derives the same corpus directly from YouTube. Full logic in scripts/yt_scrapper.py and scripts/scrape_dataset.py. Rest of the project: StegoReels.

Licensing Information

Manifest (ids, categorization, splits) under CC BY 4.0. Video content not included — remains original creators' property, under YouTube's ToS.

Citation

No paper yet. Link back to StegoReels if you use this.