Bayes training — detection that learns from your own mail.

How it works

Bayes training shows archived and quarantined mail in a single combined list — training actions are available right in the row. Every mail you classify sharpens the statistical model: MailGuard learns what your legitimate communication looks like and what your spam looks like. Above the table, live metrics show the state of the corpus (Bayes ham, Bayes spam, neural samples, scanned and learned messages).

Six classes instead of just spam/ham

Instead of only "good" and "bad", MailGuard distinguishes six training classes — so the filter learns more precisely and makes fewer mistakes on legitimate bulk mail:

• spam — unwanted advertising mail (standard spam training)
• ham — legitimate mail (standard ham training)
• phishing — train phishing attempts specifically as a pattern
• bec — business email compromise: CEO fraud and targeted impersonation
• newsletter — classify legitimate bulk mail correctly
• transactional — order confirmations, system and transactional mail

For spam and ham there are quick buttons right in the row; the four additional classes are trained via the action menu.

Bayes and neural network

Alongside the classic Bayes classifier, MailGuard trains a neural network. It kicks in from 1,000 samples per class (rspamd default behaviour) and adds learned patterns to the statistical detection that individual features alone can't capture.

Trained on your own corpus — not in someone else's cloud

The training corpus draws on two local sources: archived mail from the archive and quarantined mail from the Postfix hold queue. A full-text search over sender, recipient or subject helps you find specific examples. Crucially: the model learns from your mail on your server — there is no vendor-wide shared corpus your communication feeds into.

A filter that learns your mail reality?

Six classes, a neural network, your own corpus — adaptive detection without cloud data flow.

See pricing