telegramatlas

Understand people.
In conversation.

Real exchanges. Shared knowledge. The context between people, preserved at extraordinary scale.

Explore the conversations
THE FULL CORPUSSource-reported figures
143.76BText messages
622MUsers
18MChannels
16.15BReactions

A corpus substantially larger by message count than major published Telegram research datasets, with the conversational structure needed to study how people interact.

Start with a conversation.

1,000 sample conversations · 34 languages

Opening the collection…

BEYOND WEB PAGES

Millions of communities.
Everyday human context.

Across millions of Telegram channels and groups, people exchange news, coordinate local efforts, solve technical problems, and share files.

These conversations can reveal knowledge and perspectives that conventional website collections miss—particularly in markets where traditional media is limited. The value is in the context: who responds, what gets shared, and how a community reacts.

About this conversation

About the data

The proposal combines a reported 143.76 billion text messages with indexing by user and group. The sample lets you inspect the connections: who responds to whom, how people share knowledge, and how a group reacts.

1,000 sessions from public Telegram groups, with 153,066 messages across 34 dominant session languages.

38,149Pseudonymous people
84,963Reactions

Participant names are replaced with labels. The source reports masking identifiers in message text.

About these figures

Full-corpus totals reported in the source proposal ↗; not independently audited. Text messages exclude attachments. Approximate token counts: 13.5T (GPT-2), 9.4T (GPT-4), 7.1T (GPT-5.1).

Dataset comparisons

The proposed corpus is substantially larger by reported message count than these published Telegram research datasets.

CorpusReported messagesScope & access
This proposal ↗143.76BText messages; proposed full-corpus delivery
TeraGram (2026) ↗Over 5.9BTelegram groups and channels; text available to researchers on request
USC election dataset ↗~1.03BElection-related Telegram collection; public distilled release

Different collection scopes; raw message counts are not counts of unique conversations or training-ready examples. Access and licensing differ. Proposal totals are not independently audited. Comparison reviewed September 2026.

Sample notes

Sessions were selected for conversation quality, not randomly sampled. Summaries and language labels come from the supplied curation and may be imperfect. Reactions are not reliable sentiment labels. Media files are not included.

Participant labels such as P1 apply within one session. Stable pseudonymous IDs are retained in the underlying data. The sample spans January 2023 to April 2026.

Read right here in English

Machine translations are a reading aid. Original messages, speakers, replies and reactions stay available.