Robert Swan wrote: >How do I clear, or unlearn the bayes filter it seems that it is picking >up wrong. E-mail that is SPAM has autolearn=ham in the header and this >is wrong. > > > Is it?
The autolearner uses the score the message would have gotten if bayes was disabled, all userconf (ie: white/blacklist) rules were disabled, and the AWL was disabled. Post a X-Spam-Status header for the message in question and we can give you some more specific advice, but just because the final score indicated spam it doesn't mean the autolearner can't decide it's ham. This is particularly true for message that got heavily hit on a blacklist or AWL rule. IMHO, the default ham learning threshold in current versions of SA is begging for problems like this. I keep mine set at a tiny negative score, but also have a collection of nonspam rules with tiny negative scores. This way, autolearning as ham must be earned by hitting one of the negative scoring rules, but the negative scoring rules can't be abused by spammers as they collectively add up to less than -1.0.
