...

Again, it's important to do both.

How to train Bayes without logging on

(DanKohn)

If you don't read your mail on the account where SpamAssassin is running, it can be challenging to do mistake-based training, where you learn false negatives (i.e., spam that was not caught) as spam. One approach is redirect your false negatives and use procmail to train on them, as described in ProcmailToForwardMail.

Training *and* reporting

(KurtYoder)

...