On Thu, 12 Sep 2013 22:09:42 +0200 Mathieu R. wrote:
22:01 root@effraie01 ~ # sa-learn --username=vmail --backup v 3 db_version # this must be the first line!!! v 0 num_spam v 0 num_nonspam
sa-learn --username sets the virtual user not the unix user. (BTW sa-learn --dump magic is a quicker way of reading the metadata)
By default SA stores the bayes database files under a user's home directory. If you run sa-learn as vmail, which doesn't have a home directory, it will probably just give up.
What you need to do is set bayes_path (in local.cf) to a directory to which vmail has access, then run sa-learn as vmail. Alternately you can setup one of the SQL backends.
I'm posting this through gmane as I'm not subscribed to dovecot list. I replied in that SpamAssassin list before spotting that it was cross-posted. You can ignore most of it now, but I'll quote what I wrote about learning ham:
"I'm sceptical that the Antispam plugin can learn enough ham this way. As I understand it the only mail that gets learnt as ham will be false-positives based on the overall spamassassin score, irrespective of the Bayes result. Bayes needs (by default) 200 spams and hams to even start classifying and much more for optimal results - I don't expect to get 200 FPs in the rest of my life. Unless this is high volume server with a shared database, I'd suggest either learning a few thousand hams manually, or implementing an unsure folder. You can also mitigate the problem by autotraining with a high ham threshold, but then you really need to be careful to move all spam to the spam folder. "