mxmprhd/youtube-shorts-dataset

YouTube Shorts Dataset

1077 vertical YouTube Shorts from 48 creators across 19 categories. H.264/AAC, 5–60s, mean 37.3s.

Metadata only, no video files. Every row is a pointer to a public YouTube video. Download them with the bundled scripts, see Getting the videos.

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 19 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 usual Hugging Face convention.

Splits

Assigned per channel, not per video: a category with 1 channel keeps it in train; 2 channels → train + validation; 3+ → train, validation, test. Every category holds at least one creator out of training, so validation/test measure generalization to unseen creators rather than memorized style. Validation/test channels get 35% of the train clip quota.

category	train	validation	test	total
animals_pets	33	14	0	47
animation_edu	40	14	0	54
asmr	35	14	14	63
beauty_fashion	40	14	0	54
challenge_stunts	40	14	0	54
comedy_sketch	40	14	14	68
cooking_food	40	14	14	68
dance	40	14	0	54
diy_crafts	40	14	14	68
education_howto	40	14	14	68
fitness	24	14	14	52
gaming	5	14	0	19
high_speed_science	40	13	14	67
music_performance	25	14	0	39
nature_wildlife	40	14	0	54
sports_trick_shots	40	14	14	68
tech_review	40	14	14	68
travel_adventure	40	14	4	58
vfx_magic	40	14	0	54

test = 0 means the category has only two channels, per the rule above. gaming is thin because that channel has few vertical Shorts.

Getting the videos

pip install yt-dlp
python scripts/scrape_dataset.py \
  --channels-file config/channels.csv \
  --output-dir raw \
  --target-total 2000

Downloads to raw/<video_id>.mp4 plus a .meta.json sidecar each, and writes raw/manifest.jsonl at the end.

Notes:

• --target-total is a budget cap, not a target. The per-channel quota is ceil(target_total / n_channels). 2000 is what this corpus was built with; passing 1077 yields a smaller one.
• Resumable. raw/.yt_dlp_archive.txt records every id already tried, so re-running fetches only the shortfall.
• YouTube rate-limits long runs: past ~500 clips in one session requests start failing with This video is not available. It clears in a few hours. --cookies-from-browser chrome and a larger --sleep-interval push further in one go.
• Expect slightly fewer clips than listed here. Videos get deleted and channels go private over time.

Caveats

• Channels were picked by hand for content diversity, not sampled randomly. Diverse, not representative.
• category is per channel (the creator's primary content type), not per video, so a clip may not match its category's subject.
• View/like counts are a snapshot from scrape time.
• Deduplicated by video id only. No perceptual near-duplicate detection.

License

The manifest (ids, categorization, splits) is CC BY 4.0. The videos are not included and remain their creators' property, under YouTube's Terms of Service.