Upload a Telegram export of your chat and we train a personal ML layer on how your community actually talks. It vetoes false spam alarms on your group's normal speech — it never bans on its own. It is trained on anonymised data, checked before it goes live, and can be toggled off any time. Available on Pro and above with a monthly import quota; manage it in settings.
Want to set this up? Do it right in your Telm account.
Open in your dashboard1What the personal model is
Our global spam model is trained on spam from thousands of groups. But every community has its own way of talking — jargon, slang, in-jokes — and occasionally the global model mistakes that normal speech for spam.
The personal model fixes this. You upload a Telegram export of your own chat, we train an additional model layer on it, and it learns what normal looks like specifically in your group. From then on it can veto false spam alarms that do not fit your community's voice.
Crucially, it only ever removes false alarms — it can never ban anyone by itself. The worst it can do is let a message through for a human to look at.
2How it helps
The personal model sits alongside the global one. When the global model flags a message as spam but the personal model recognises it as your community's normal speech, the spam verdict is held back and the message is sent to review instead of being removed.
The result is fewer false positives on the topics and phrasing that are perfectly normal in your group, without weakening protection against real spam.
- Best for communities with a distinctive vocabulary — trading desks, dev chats, fandoms, regional slang.
- It works in real time, adding no noticeable delay.
- It changes only the false-alarm side; genuine spam is still caught by every other layer.
3How training works
You do not have to prepare anything technical. You export your chat history as a file, upload it, and the rest happens automatically while you watch the progress.
Behind the scenes your export is cleaned and anonymised, a model is trained on your community's speech, and it is checked before it is ever switched on.
- Your export is anonymised first — usernames, phone numbers, emails and links are replaced with placeholders before anything is trained.
- It learns from your trusted regulars, so it captures your community's genuine voice.
- A model is trained on your community's speech and then tested for accuracy.
- Only a model that passes every check is published and applied; a weaker one is discarded.
4Checks that keep it safe
A personal model is only switched on if it passes automatic checks. If it fails any of them, it is rejected and never applied, so it can never make your protection worse.
- It must not let real spam through — it is checked against known spam.
- It must recognise your group's own speech — tested on a sample of your trusted members.
- It must handle clean messages without mistakes.
- All uploaded data is anonymised (usernames, phones, emails and links are replaced with placeholders) before training.
5Privacy and your data
The export never becomes public and is not kept around. It is used to train your model and then removed.
See what data Telm sees and data privacy and storage for the full picture.
- Personal details are replaced with placeholders before any training happens.
- The raw uploaded file is deleted after processing; abandoned uploads are cleaned up automatically.
- The trained model captures speech patterns, not readable message logs.
6Uploading your chat export
You export your group's history from Telegram Desktop as JSON, then upload it through the import wizard in the dashboard. Processing runs automatically and you can watch its progress.
The larger and cleaner the export, the better the model learns your community — a bigger history also earns a longer subscription bonus (see below).
7Plans, quotas and bonuses
Imports are limited per month, and the limit depends on your plan.
- Pro — 1 import per month per group.
- Business — 2 imports per month per group.
- Organizations — up to 4 imports per month per group.
- Bonus: a successful import adds 7 days to your subscription, or 14 days for an export of 10,000+ messages.
- Rejected uploads for a bad file or too few messages do not use up your quota.
8Managing the model
Once trained, the personal model is applied automatically. You stay in control of it from the group settings.
There is a per-group toggle to turn the personal model on or off — turning it off does not delete the trained model, so you can re-enable it later without re-importing.
- Toggle it on or off per group at any time.
- Turning it off keeps the trained model, so re-enabling needs no new import.
- Re-import within your monthly quota if your community's style changes a lot.