[Rspamd-Users] Per user bayes filtering

Mark Constable markc at renta.net
Sat May 9 01:59:11 UTC 2020

On 9/5/20 1:44 am, Stefan Bauer wrote:
> Per user means per user. Our observations have shown, that you need
> a decent amount of mails trained by each user to have a real base to
> make decicions.

Thank you. That's the first positive reference I've seen confirming
that per user works and actually means just that. So does this mean
when I train the bayes filter then I should do something like?...

rspamc learn_spam --user jo at blo.com

and should I tweak my sieve scripts to look for some $USER env var
and make sure the users retraining actions also include --user ?

> thats why we do not have per user. As users are unable to
> differentiate between spam an unwanted mails, i do not advice you to
> let your users train filters.

But if per user bayes filtering does not affect global stats for
other users then why would it matter if one user trained unwanted
promotional material as spam if it won't affect other users stats?


As a long term spamprobe user (with a separate filtering database per
user) we found that if some folks trained promotional material as
spam then, fine, it works for them as it gets rid of "junk", ie; stuff
they don't want in their inbox. My admin@ Junk folder had about 300
messages for the previous 7 days and I was getting about 1 spam per
week through to my Inbox, after a few years of training.

My naive and uneducated change to rspamd meant a lot of our customers
started getting inconsistent spam/ham results for the last 2 weeks
and are generally not happy.

Mark Constable
0419 530 037

Mark Constable
0419 530 037

More information about the Users mailing list