public wall posts, comments, discussion boards, photos, videos, communities and profiles from VK (VKontakte), scraped continuously since October 2026. posts, comments, topics, photos and videos are uploaded every day, and the communities and users snapshots weekly.

from datasets import load_dataset

posts = load_dataset("tiagozip/vk", "posts", split="train", streaming=True)
communities = load_dataset("tiagozip/vk", "communities", split="train")

format

config	notes
posts	split by scrape day, `data/posts/fetched=YYYY-MM-DD/*.parquet`
comments	split by scrape day, `data/comments/fetched=YYYY-MM-DD/*.parquet`
topics	discussion board threads, `data/topics/fetched=YYYY-MM-DD/*.parquet`
topic_comments	posts inside those threads, `data/topic_comments/fetched=YYYY-MM-DD/*.parquet`
photos	split by scrape day, `data/photos/fetched=YYYY-MM-DD/*.parquet`
videos	split by scrape day, `data/videos/fetched=YYYY-MM-DD/*.parquet`
communities	snapshot, one row per community
users	snapshot, one row per user seen

everything except the snapshots is split by the day it was scraped (fetched_at), not by when it was posted. the scraper revisits walls, so the same post can appear in more than one day's folder with newer counters. to get one row each, dedupe on the id columns and keep the row with the newest fetched_at. the id columns are (owner_id, post_id) for posts, (owner_id, post_id, comment_id) for comments, (owner_id, topic_id) for topics, (owner_id, topic_id, comment_id) for topic_comments, (owner_id, photo_id) for photos and (owner_id, video_id) for videos. inside the a-* and b-* files of a single day, a row can also repeat, and the same rule applies.

ids follow VK's convention: a community owner is a negative id, a user is positive.

posts

column	type	notes
`owner_id`	int64	wall owner, negative for communities
`post_id`	int64	id within the owner
`from_id`	int64	author, may differ from owner
`date`	timestamp
`text`	string
`comments`	uint32	counter at fetch time, not current
`likes`	uint32	same
`reposts`	uint32	same
`views`	uint32	same
`is_pinned`	bool
`post_type`	string	post, copy, reply, ...
`signer_id`	int64	0 unless signed by a specific user
`copy_owner_id`	int64	repost source owner, 0 when not a repost
`copy_post_id`	int64	repost source post id
`attachments`	string	raw JSON array (photos, videos, polls, docs)
`fetched_at`	timestamp

comments

owner_id, post_id, comment_id, from_id, date, text, likes, reply_to_user, reply_to_comment, thread_count, attachments (raw JSON), fetched_at.

topics

owner_id, topic_id, title, created, created_by, updated, comments, is_closed, fetched_at.

topic_comments

owner_id, topic_id, comment_id, from_id, date, text, likes, attachments (raw JSON), fetched_at.

photos

owner_id, photo_id, album_id, user_id, date, text, url, width, height, likes, comments, fetched_at. url points at VK's image server, the images themselves are not included.

videos

owner_id, video_id, title, description, date, duration (seconds), views, comments, likes, is_clip, player (embed url), fetched_at. the video files are not included.

communities

id, name, screen_name, type, is_closed, members_count, description, activity, verified, deactivated, wall, fetched_at.

users

id, first_name, last_name, screen_name, is_closed, deactivated, sex, city, country, followers_count, last_seen, fetched_at.

what this is not

not complete. VK has on the order of tens of billions of public posts, so consider this a large and growing sample. communities with more members and more recent posts are the most complete.
engagement counts are snapshots. every counter is whatever it read at fetched_at. a post fetched moments after it was stored will show near-zero engagement. do not treat these as final values.
private and closed walls are absent. only content readable without logging in is here.

license

tiago.zip/license

disclaimers

access can be revoked at any time and all data is provided as-is. this dataset contains personal information about real people, a lot of whom are not public figures. this dataset is not a reflection of the views of the author or the institutions they are affiliated with. please be respectful. every post in this dataset was publicly readable at the time it was fetched, and all scraping was done unauthenticated.

VK and VKontakte are trademarks of their respective owners. this dataset is independent and unaffiliated.

opting out

you may opt out at optout.tiago.zip.

citation

@misc{vk,
  title  = {vk: a corpus of public VKontakte posts},
  author = {tiagozip},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tiagozip/vk}
}

Downloads last month: 345

Total file size:

763 GB