I understand SpamBayes works best when trained on roughly equal numbers of ham & spam messages. After initial training, I've been training it only on messages it got wrong or was unsure about. Should I also be including some (or all) correctly-labeled ham & spam in the training? Thanks.