Just thought I'd share in case anyone else ends up stumped by this problem...
Zimbra's bayes scoring has been atrocious for us since day one, especially compared to our standalone amavis system. I think I finally found that the reason is that our users are a little bit brain dead. I've clicked View Mail on the spam account and discovered that less than 1 in 10 messages are what I would consider actual spam. Most of the rest includes things like messages from our internal organization-wide mailing lists, forwards from their friends, and simply information they didn't want. As you can imagine, this seems to be confusing the hell out of spamassassin/bayes, and it's actually negatively scoring a lot of spam, letting it through the filter even if it would have otherwise have been tagged just by rules.
I'm curious if anyone has had any luck overcoming this problem. It might get me close enough to fixing the problem if I could put filters on the spam account that would delete messages from local addresses before zimbra trains on the account, but since the spam is included as an attachment, that won't really work. For now I've commented out the zmtrainsa line in the zimbra crontab and just manually run zmtrainsa on a few trusted mailboxes.


LinkBack URL
About LinkBacks

