The Bayesian filter "learns" about spam by seeing what the other filters do. The more accurate the other filters are, the more accurate the bayesian filter becomes.
Usually the other filters, like our SFDB database, MX tests, SPF, RBL etc will provide excellent results. Also important during the training period is to go thru the quarantined emails, and to force the delivery of valid emails that were incorrectly blocked. This will also tell SpamFilter what emails were blocked by mistake, allowing it to self-adapt.

Please note that Bayesian filter were very effective a few years ago, but nowdays spammers have learned to fool them more easily, rendering them less effective.

We confirm. The Bayesian filter will learn even if you're not quarantining the emails.