Skip to main content
Add to group
Community-Trained AI

Personal
ML model

A model that speaks your group's language. Train it on your chat export and Telm learns how your community talks — vetoing false alarms even at 97% confidence.

  • Trained on your chat
  • Never bans on its own
  • Updates weekly
“Wen moon ser? Aping the presale, wagmi 🚀”
Global model
Spam — 97% confidence
Your personal model
Normal speech in this group
Verdict vetoed — sent to review, not deleted

From chat export to personal model

1

Upload your chat export

Export your group's chat history and upload it. That's the whole setup — no labeling, no configuration.

2

Telm learns your community's voice

The model trains on your group's real messages and learns what normal looks like specifically in your chat — jargon, slang, in-jokes included.

3

It stays current

The model updates weekly, and you can upload a fresh export any time so it keeps up with how your community talks now.

Designed to never overreach

Veto power, nothing more

The personal model can only overturn false spam alarms — even at 97% confidence. It never bans, deletes, or punishes on its own.

Vetoed verdicts go to review

When it overrules the global model, the message is sent for review instead of being removed — your members don't get silenced by mistake.

Real spam is still caught

It only touches the false-alarm side. Genuine spam is still stopped by blocklists, rules, the global model, and the LLM.

Quality-checked before it goes live

A trained model is enabled only if it actually improves detection for your group. If it doesn't, it never touches a verdict.

Why a global model isn't enough

Telm's global spam model is trained on spam from thousands of groups, and that breadth is its strength — and its blind spot. Every community has its own way of talking: jargon, slang, in-jokes, phrasing that looks suspicious everywhere except in your chat. The personal model closes that gap. You upload your group's chat export, and Telm trains an additional model layer on how your community actually talks. From then on it sits alongside the global model and can veto a false alarm even at 97% confidence — when a flagged message reads like your group's normal speech, the verdict is held back and sent to review instead of the message being removed.

The design is deliberately one-sided: the personal model never bans, deletes, or punishes on its own. It can only step in on the false-alarm side, which means the worst it can ever do is let a message wait for a human — real spam is still caught by every other layer of the pipeline. The model updates weekly, and you can upload a newer export whenever you want it to keep pace with how your community talks now. The payoff is simple: fewer false positives on the topics and phrasing that are perfectly normal in your group, with none of the risk of a custom model gone rogue.

Frequently asked questions

What do I need to train the model?

A chat export of your group. Upload it and Telm trains a personal model layer on how your community actually talks — no technical preparation needed.

Can the personal model ban or delete anything on its own?

Never. It's veto-only by design: it can overturn a false spam alarm, and the overturned message goes to review. The worst it can ever do is make a message wait for a human.

What if the global model is very confident?

The personal model can still step in — it vetoes false alarms even at 97% confidence when the flagged message reads like your community's normal speech.

Does it weaken my spam protection?

No. It only changes the false-alarm side of the equation. Real spam is still caught by every other layer — blocklists, 150+ rules, the global model, and the LLM for borderline cases.

How does it keep up as my community changes?

The model updates weekly, and you can upload a newer chat export whenever you like to keep it in sync with how your group talks today.

Ready to Get Started?

Start protecting your community today. Free plan available.