We have a script that does some unattended spam learning from a designated spam box that our users report spam to. I ran a dump on the bayes db and I noticed some tokens that shouldn't be ranked so highly, so I started watching the spam box and it seems that some users got "confused" and reported ham as spam and so spamassassin has learned these tokens incorrectly... Is there any way to just delete these incorrect tokens and leave the others?
Not easily.. You'd have to dump the whole db as text, edit it, and then reload it.
A bit easier would be to re-train the mis-learned messages as the correct type. SA will recognize that the message was previously mis-learned and compensate appropriately.
