brianhliou/youtube-comment-exposure

YouTube Comment Exposure Corpus v1

1,345,353 top-level YouTube comments from 633 fully-enumerated videos across 34 channels, with the field the existing public corpora leave out: when each comment was posted.

Why this exists

The two large public YouTube comment datasets, YT-30M and YTCommentVerse (~32M comments between them), carry like counts with no timestamps and no reply counts. A like count without a post time is close to unusable for anything causal, because a comment posted in a video's first hour gets far more impressions than one posted a week later. Any model trained on those corpora is fitting that head start without being able to see it.

This corpus records post time, fetch time, reply count, video statistics, and the position YouTube's own "Top comments" ranking gave each comment.

Files

file	rows	contents
comments.jsonl.gz	1,345,353	one top-level comment per line
videos.jsonl.gz	633	video metadata and collection counts
manifest.json		row counts, field lists, SHA-256 per file

Comment fields

comment_id, video_id, author_hash, text, like_count, reply_count, published_at, fetched_at, age_hours (comment publish minus video publish), relevance_rank (position on the displayed top page, null if absent).

Video fields

video_id, channel_handle, published_at, title, category_id, duration_seconds, is_short, view_count, like_count, comment_count, stats_fetched_at, collected_comments.

How it was collected

Channels were picked to span science, tech, gaming, commentary, music, geopolitics and cooking, then filtered to videos published between 2025-02-13 and 2026-07-14 with 200–20,000 reported comments. That window leaves at least 30 days for like counts to settle before collection.

Every video is enumerated completely. commentThreads.list returns newest-first under order=time, so a per-video page cap would drop the oldest comments, which are exactly the high-exposure ones. Videos too large to enumerate inside the API's daily quota were excluded when the population was frozen, and the counts are recorded in the population manifest rather than dropped silently.

Collection ran 2026-08-13 to 2026-08-15 against the YouTube Data API v3. All 633 videos completed; none were skipped or failed.

What's in the corpus, in aggregate

82.1% of comments have zero likes and 91.9% have one or fewer. Mean log1p(likes) by within-video comment-age decile runs 0.721 down to 0.133, monotonically. Comments appearing on the displayed top page average 419.5 likes against 1.13 for those that don't.

Known limits

Top-level comments only. Replies sit under their parent rather than competing for the video's ranking, so they're a different exposure regime. reply_count is kept as a covariate; reply text isn't collected.

The comment-count band is a real restriction. Videos above 20,000 comments are absent, and those are where exposure effects are largest. Findings shouldn't be extrapolated to the largest videos on the platform.

Like counts are a single observation. fetched_at records when, and a comment collected on day one of the run has had less settling time than one collected on day three.

English-dominant, creator-skewed. 34 channels is enough for a held-out-channel split and nowhere near a representative sample of YouTube.

Collection order was smallest-video-first, so any partial copy of this corpus is size-truncated rather than a random subsample.

Author identifiers

author_hash is a SHA-256 of the author's channel ID with a random salt that was generated at export and never written down. That prevents linking an author between this dataset and any other one.

It does not make the data anonymous. Comment text is public, indexed, and searchable, so anyone determined can recover an author from the text. Treat this as pseudonymized public content, not as de-identified data.

Licensing and terms

Comment text belongs to the people who wrote it. This corpus is released for research under CC BY 4.0 as applied to the collection, the schema, and the derived fields, which follows the precedent set by YTCommentVerse.

YouTube's API Terms of Service restrict storage and redistribution of API data, and this release is in the same grey area every public YouTube comment corpus occupies. Anyone reusing it should read those terms and reach their own conclusion. Takedown requests for specific comments will be honored.

Citation

@misc{liou2026commentexposure,
  title  = {YouTube Comment Exposure: 1,345,353 Comments with Post Timestamps},
  author = {Liou, Brian H.},
  year   = {2026},
  note   = {https://brianhliou.com/publications/youtube-comment-exposure/}
}

The paper reports what the timestamps make measurable: exposure accounts for 83.5% of the ordering skill available above chance, against 8.2% for comment content.

Reproducing it

Collector, frozen population manifest, and analysis code are at github.com/brianhliou/youtube-comment-exposure under MIT. The population manifest pins the exact 633 video IDs and the filters that selected them, so the sample can be rebuilt or extended without guessing at the selection rules.