Datasetpublic

starnikovoleg/telegram-public-channels-2026-W36

Telegram public channels: a 5,177-channel snapshot with topics and a recommendation graph A single snapshot of 5,177 public Telegram channels, measured on 31 August 2026, together with the recommendation graph Telegram itself exposes between them. Public Telegram data is hard to get in tabular form. The libraries that read it need a phone number and a user session, and the two Telegram datasets that rank on Kaggle today are both from 2021. This is a current measurement… See the full description on the dataset page: https://huggingface.co/datasets/starnikovoleg/telegram-public-channels-2026-W36.

sourceHugging Facecc-by-4.0updated 10d agoView on Hugging Face

0likes112downloads

Dataset Card

Telegram public channels: a 5,177-channel snapshot with topics and a recommendation graph

A single snapshot of 5,177 public Telegram channels, measured on 31 August 2026, together with the recommendation graph Telegram itself exposes between them.

Public Telegram data is hard to get in tabular form. The libraries that read it need a phone number and a user session, and the two Telegram datasets that rank on Kaggle today are both from 2021. This is a current measurement, published as plain CSV.

What is inside

File	Rows	What it holds
`channels.csv`	5,177	one row per channel: username, title, topic, subscribers, posts per day, verified flag, creation date, how many channels recommend it
`recommendations.csv`	10,911	edges of the "similar channels" graph: source channel, recommended channel
`topics.csv`	20	per topic: channel count, total and median subscribers, average posting rate, verified count
`weekly-index.csv`	1	the week-level summary row for this snapshot

Totals worth knowing before you load it: 645,049,485 subscribers across all channels, 20 topics, 561 verified channels. posts_per_day is filled for 5,170 channels; for 682 of those it is a lower bound and carries the posts_per_day_is_lower_bound flag.

The graph, precisely

As measured on 17 September 2026: 10,911 edges, 1,154 source channels, 5,244 distinct nodes. 274 recommended channels have no row in channels.csv — 272 of them sit below the 1,000-subscriber threshold this dataset publishes at, and 2 were not part of the crawl. If you join the two files on username, that is the difference you will see, and now you know where it comes from.

What this dataset does not contain

—No growth. Growth needs two measurements and this is one. The weekly_growth_pct and growth_days columns exist in the schema and are empty on purpose, rather than filled with guesses.
—No private data. Public channels only: no private chats, no closed channels, no user identities, no message text of individuals.
—No promise of weekly updates yet. A second week is being measured. When it lands, this becomes a series and the growth columns fill themselves.

Ideas to start with

—Which topics concentrate audience and which spread it thin: the median against the mean per topic tells that story in one chart. In this snapshot the ratio runs from 1.3 in startups to 22.4 in regional-es.
—Communities in the recommendation graph: it splits into 60 connected components, the largest holding 3,606 nodes, and several islands are single-language or single-purpose.
—Posting rate against audience size: the correlation is 0.019 across the 4,488 rows where the rate is exact rather than a lower bound. Weaker than most people assume.

Provenance

Measured by the tgAtlas crawler on 31 August 2026 (week 2026-W36) through Telegram's own MTProto API, then aggregated. The same data powers the public pages at tgatlas.org and is served as open CSV at tgatlas.org/data.

Attribution: tgAtlas, with a link to the page the number came from.