Hi, > Instead, be more selective about the spam you train. Only train > messages that completely missed with respect to Bayes (e.g. a spam > that got BAYES_00 or a ham that got BAYES_80) rather than corner cases > (e.g. a spam with BAYES_50 that got marked as spam, a spam that got > marked as BAYES_80). Try to train as much inbound ham as possible > (but again, not internal messages that never hit the live internet). > If you use autolearn, bump the bayes_auto_learn_threshold_spam up some.
This sounds like really good advice. I've recently enabled autolearn and I'm a bit concerned that my database is skewed. I found quite a few spams in the quarantine with a bayes score less than 50. However, I'm not sure that wasn't the case before I started the autolearn. In either case, is there a way to exclude mails with USER_IN_WHITELIST altogether? I have my ham level set at -0.3, but the USER_IN_WHITELIST (and there are quite a few) adds -100.0, automatically making it ham. I'm concerned that a spoofed mail passing through the whitelist could skew the db without my knowledge and without my ability to control it. Thanks, Alex
