Kairos Dataset · ITMO University · Lab 1

Who engages with a post, and why? Five thousand feeds to find out.

LinkedIn authors, the feeds they were served and every media file of their posts, with a label for how each post did against its author's usual level. The text is pseudonymised; the media are not.

Data sponsored by SOMIN.

What is inside

One release, two archives, one label per post.

Authors

5,433

Profiles with their feeds, coded as AUTHOR_xxxx.

Posts

151,351

Feeds as served, reposts included.

Labelled

89,018

Posts with the author's own counters and five earlier posts.

Media files

161,589

Images, videos, PDFs, pages, avatars and logos - each stored once.

The task

Did the post do better or worse than its author usually does?

Below

Under ×0.577

Reactions well under the median of the author's earlier posts. ~20 %.

Typical

Between

Within the author's usual range. ~60 %.

Above

Over ×3.182

Reactions well above the author's usual level. ~20 %.

Split

80 / 20 by author, seed 42.

Train: 3,295 authors, 71,395 labelled posts. Test: 824 authors, 17,623 posts. All posts of one author sit on one side. The cutoffs are fitted on train and applied unchanged to test.

Plain reposts carry the original's counters, so they have no label and do not count towards the level.

Download

Sixteen volumes, one password, sent to you by e-mail.

Volume	Size	SHA-256

core (5 volumes, 50.5 GB): text, labels, manifest, images and PDFs. video (11 volumes, 114.1 GB): the post videos. You can work without the video archive.

Each volume is too large for Google Drive to scan for viruses, so a link opens a warning page first: press Download anyway. Check the file against its SHA-256 afterwards.

Request the password

Tell us who you are.

The volumes are encrypted. Fill this in and the password arrives at the address you give.

Getting started

From volumes to a training table in four steps.

Download every volume of a set into one folder.

Check each file against the SHA-256 above:

certutil -hashfile lab1_core.7z.001 SHA256   # Windows
sha256sum lab1_core.7z.*                    # Linux / macOS

Unpack by opening the first volume with 7-Zip (Windows), Keka (macOS) or 7z (Linux, package p7zip-full); the others are picked up automatically:
```
7z x lab1_core.7z.001
```
Join labels, posts and media - README.md in the archive has the code.

Layout

authors/AUTHOR_xxxx.json    profiles
posts/AUTHOR_xxxx.json      feeds, urn = post code
media/<ab>/<sha256>.<ext>  media, once per file
media/manifest.jsonl        post → file
labels.csv                  author, post, split, label
split.json                  authors and cutoffs

Rules of use

The text is coded. The pictures, videos and PDFs are not.

Pseudonymised

Text only

People, organisations, handles, e-mails, phones and links in the text are replaced by codes. Faces, slide names, logos and voices in the media are left as they are.

Do not

Re-identify anyone

Do not try to link a code to a real person, and do not publish examples that would let others do it.

Keep it

Inside the course

Use it for the lab, your project or research. Do not re-host the data or pass on the password.

SOURCE Public LinkedIn profiles and feeds, collected and sponsored by SOMIN in 2026 for the ITMO course Social Media Analytics, Agents & LLMs.