ScriptSmith/sponsorblock-youtube-metadata-2024

SponsorBlock YouTube Metadata Dataset

A dataset of YouTube video metadata collected from a subset of videos in the SponsorBlock database. This dataset contains metadata, subtitles, engagement heatmaps, live chat, and channel playlist information for popular YouTube videos.

Quick Stats

Metric	Value
Total videos	154,536
Videos with subtitles	62,819 (41%)
Videos with heatmaps	~92%
Videos with chapters	~32%
Livestream recordings	4,300
Unique channels	20,061
Channel playlist entries	~20.8M
Average video duration	28.6 minutes
Average view count	1.6M views

Collection Period

Data Type	Collection Date	Content Date Range
Video metadata	Aug 9 - Sep 30, 2025	May 2019 - Dec 2024
Channel playlists	Oct 2-5, 2025	Varies by channel

Note: All metrics (view counts, likes, subscriber counts) are snapshots from the collection date and do not represent historical data.

Data Source

Videos were selected from a subset of the SponsorBlock database, a crowdsourced browser extension for skipping sponsor segments. The subset was created by sorting all videos by their SponsorBlock vote count (total upvotes on submitted segments) and selecting the top ~167K video IDs. This selection method biases the dataset toward:

• Highly-engaged SponsorBlock content — videos with many user-submitted and upvoted segments
• Popular videos with large, active communities
• Gaming, tech, and entertainment content (common SponsorBlock categories)
• Videos with sponsor segments (brand deals, mid-roll ads)
• English-language content primarily, with significant Russian, Hindi, and European language representation

Note: This dataset does not represent a random sample of YouTube or even a random sample of SponsorBlock. Videos were specifically selected for having high engagement within the SponsorBlock community.

Category Distribution

Category	Percentage
Gaming	15%
Entertainment	11%
People & Blogs	8%
Science & Technology	6%
Education	5%
Other	55%

Language Distribution

Language	Percentage
English (en, en-US)	25%
Russian	16%
Hindi	2.5%
Polish	2%
Other	54.5%

Dataset Configurations

video_metadata

Full metadata for 154,536 individual YouTube videos.

Column	Type	Description
id	string	YouTube video ID (11 characters)
title	string	Video title
fulltitle	string	Full video title
description	string	Video description (may be truncated)
upload_date	string	Upload date (YYYYMMDD format)
timestamp	int64	Unix timestamp of upload
duration	int64	Duration in seconds
view_count	int64	Number of views at collection time
like_count	int64	Number of likes at collection time
comment_count	int64	Number of comments at collection time
channel	string	Channel display name
channel_id	string	YouTube channel ID (24 characters, starts with UC)
channel_url	string	Full channel URL
channel_follower_count	int64	Subscriber count at collection time
channel_is_verified	bool	Whether channel has verification badge
uploader	string	Uploader name (usually same as channel)
uploader_id	string	Uploader handle (@username)
width	int64	Video width in pixels
height	int64	Video height in pixels
fps	float64	Frames per second
vcodec	string	Video codec (e.g., vp9, av01, avc1)
acodec	string	Audio codec (e.g., opus, mp4a)
dynamic_range	string	SDR or HDR
aspect_ratio	float64	Width/height ratio
age_limit	int64	Age restriction (0 = none, 18 = adult)
availability	string	public, unlisted, or private
live_status	string	not_live, is_live, was_live, is_upcoming
was_live	bool	True if video was a livestream
is_live	bool	True if currently live
language	string	Detected video language (ISO 639-1)
playable_in_embed	bool	Whether embedding is allowed
categories	string	JSON array of YouTube categories
tags	string	JSON array of video tags
thumbnail	string	URL of highest quality thumbnail
chapters_json	string	JSON array of {title, start_time, end_time}

subtitles

Extracted subtitles for 62,819 videos (41% coverage).

Column	Type	Description
video_id	string	YouTube video ID
language	string	Subtitle language code
full_text	string	Concatenated plain text of all subtitles
segments_json	string	JSON array of {start, end, text} segments

Note: These are auto-generated captions from YouTube, not human-created subtitles.

channel_playlists

Video entries from 20,061 channel upload playlists (~20.8M total entries).

Column	Type	Description
id	string	YouTube video ID
title	string	Video title
description	string	Video description
duration	int64	Duration in seconds
view_count	int64	Number of views
timestamp	int64	Unix timestamp
upload_date	string	Upload date (YYYYMMDD)
channel	string	Channel name
channel_id	string	Channel ID
playlist	string	Playlist name (usually "Uploads from [Channel]")
playlist_id	string	Playlist ID
playlist_count	int64	Total videos in channel's upload playlist
playlist_index	int64	Position in playlist (1 = newest)
thumbnail	string	Thumbnail URL

Use cases: Channel growth analysis, upload frequency patterns, content strategy analysis.

heatmaps

Viewer engagement heatmaps showing which parts of videos are most watched. ~92% of videos have heatmap data, with ~100 segments per video.

Column	Type	Description
video_id	string	YouTube video ID
start_time	float64	Segment start in seconds
end_time	float64	Segment end in seconds
value	float64	Relative engagement score (0.0 - 1.0)

How to interpret: Values represent relative rewatch/skip behavior. Higher values indicate segments that viewers replay or don't skip. Values near 0 indicate frequently skipped segments (often intros, sponsor reads).

Use cases: Identifying engaging content patterns, sponsor segment detection, intro/outro detection, content quality analysis.

live_chat

Chat messages from 4,300 livestream recordings.

Column	Type	Description
video_id	string	YouTube video ID
message_id	string	Unique message ID
timestamp_usec	int64	Message timestamp in microseconds
video_offset_msec	int64	Video playback time when message appeared
author_name	string	Display name of message author
author_channel_id	string	Author's YouTube channel ID
message	string	Message text content
message_type	string	text, paid (Super Chat), membership, gift

Use cases: Chat sentiment analysis, community engagement patterns, Super Chat analysis.

Raw Data Archives

In addition to the processed parquet files, the complete raw yt-dlp output is available as compressed archives in the raw/ directory:

Archive	Contents	Compressed	Uncompressed
video_metadata.tar.zst	154K .info.json files	1.6 GB	56 GB
subtitles.tar.zst	62K .vtt files	2.5 GB	12 GB
live_chat.tar.zst	4K .live_chat.json files	897 MB	14 GB
channel_playlists.tar.zst	20K channel .json files	3.4 GB	48 GB
Total		8.2 GB	130 GB

Extracting Raw Data

# Extract with zstd
tar -I zstd -xvf raw/video_metadata.tar.zst

# Or with explicit zstd command
zstd -d raw/video_metadata.tar.zst -c | tar -xvf -

Additional Fields in Raw Data

The raw JSON files contain fields not included in the parquet files:

• formats - All available streaming URLs and quality options
• thumbnails - All thumbnail resolutions
• automatic_captions - Caption availability for 150+ languages
• subtitles - Manual subtitle availability
• requested_formats - Selected format details
• http_headers - Request headers used

Usage Examples

Basic Loading

from datasets import load_dataset

# Load video metadata
videos = load_dataset("ScriptSmith/sponsorblock-youtube-metadata-2024", "video_metadata")
print(f"Loaded {len(videos['train']):,} videos")

# Load with streaming for large configs
playlists = load_dataset(
    "ScriptSmith/sponsorblock-youtube-metadata-2024",
    "channel_playlists",
    streaming=True
)

Analyzing Engagement Patterns

import json
from datasets import load_dataset

videos = load_dataset("ScriptSmith/sponsorblock-youtube-metadata-2024", "video_metadata")
heatmaps = load_dataset("ScriptSmith/sponsorblock-youtube-metadata-2024", "heatmaps")

# Find videos with high early drop-off (bad intros)
# Join on video_id, look for low values in first 30 seconds

Working with Subtitles

import json
from datasets import load_dataset

subtitles = load_dataset("ScriptSmith/sponsorblock-youtube-metadata-2024", "subtitles")

# Parse segments
for row in subtitles['train']:
    segments = json.loads(row['segments_json'])
    for seg in segments:
        print(f"{seg['start']:.1f}s: {seg['text']}")

Channel Analysis

from datasets import load_dataset
from collections import Counter

playlists = load_dataset("ScriptSmith/sponsorblock-youtube-metadata-2024", "channel_playlists")

# Count videos per channel
channel_counts = Counter(row['channel_id'] for row in playlists['train'])
print(f"Most prolific: {channel_counts.most_common(10)}")

Known Limitations

1. Temporal snapshot: All metrics are from collection time (Aug-Oct 2025) and don't reflect current values or historical trends.

2. Selection bias: Videos come from SponsorBlock, biasing toward sponsored content and certain genres (gaming, tech).

3. Missing data: Not all videos have subtitles (41%), heatmaps (92%), or chapters (32%).

4. Deleted content: Some videos may have been deleted or made private since collection.

5. Auto-captions: Subtitles are YouTube's auto-generated captions, which may contain transcription errors.

6. Truncated descriptions: Very long descriptions may be truncated.

7. Regional availability: Some videos may be region-locked; metadata reflects availability from the collection location.

License

This dataset is released under CC-BY-4.0. The underlying video content belongs to the respective creators. This dataset contains only metadata, not the actual video/audio content.

Citation

@dataset{sponsorblock_youtube_metadata_2025,
  title={SponsorBlock YouTube Metadata Dataset},
  author={Smith, Adam},
  year={2025},
  url={https://huggingface.co/datasets/ScriptSmith/sponsorblock-youtube-metadata-2024},
  note={Metadata for 154K YouTube videos from the SponsorBlock database}
}

Acknowledgments

• SponsorBlock for the video ID database
• yt-dlp for metadata extraction
• YouTube creators whose content is represented in this dataset