Klar Spam Demo (toy model)

A deliberately small, deliberately weak spam classifier, published only so the open-source Klar milter builds and runs out-of-the-box without the commercial model.

This is not a usable spam filter. It is a 4-layer XLM-RoBERTa distilled from Klar's production classifier purely to demonstrate the engine's end-to-end pipeline (encode → 4-label head → decision). Binary spam accuracy is ~0.79 on a held-out split; the production model is far better and is licensed separately.

Labels

0=gibberish 1=marketing 2=regular 3=spam

The engine treats anything that is not regular as spam-ish.

Use

# inside the open-source engine repo:
make import-demo MODEL=icosha/klar-spam-demo
make build
./engine/build/spam_classifier ./engine/model

Licence

Proprietary — Klar Model License (LICENSE), the same licence as Klar's production model. You may download and use this toy model for evaluation and to run the engine demo; production use, redistribution, and derivatives require a commercial agreement (hello@klar.im). The engine code is separately AGPLv3.