On 18 Mar 2004 12:27:01 +0200, Deon de Villiers <[EMAIL PROTECTED]> wrote: > >Do you reckon my theory on autolearning lots of ham causing lots of spam >to start getting thru sounds feasible (as bizarre as it sounds while I >read it back to myself)?
It sounds like autolearning *at all* might be your problem. Unless you very that the *none* of the messages are autolearned incorrectly, you'll be better off manually training only with hand verified spam and ham. Hope this helps, Alan >Thanks >Deon. > >On Wed, 2004-03-17 at 04:39, Robert Menschel wrote: >> Hello Deon, >> >> Tuesday, March 16, 2004, 4:26:48 AM, you wrote: >> >> DdV> Is it simply a matter of playing around with the autolearn >> DdV> thresholds until I find a good balance? >> >> Or better: playing around with the autolearn thresholds until you find >> the ones that work for you (balance not necessary). >> >> DdV> I would dearly like to utilise autolearning as I am running SA in a web >> DdV> hosting environment and so have numerous servers hosting even more >> DdV> numerous domains all talking to the same server via spamc/spamd and >> DdV> hence utilising the same bayesian database. If I can get autolearning >> DdV> working then this seems to be the right way forward for me. >> >> I use a non-spam threshold of -2 and a spam threshold of 20. >> >> The non-spam of -2 ensures that spam which sneaks by hitting no rules >> does not get wrongly learned as ham. I'm very happy with that setting. Of >> course, it requires a number of negative-scoring rules in your local >> configuration to work. >> >> The spam threshold of +20 is probably a bit high. It catches a goodly >> amount of spam, but could catch more if I lowered it. Since I manually >> feed ALL spam and ham into sa-learn, auto-learned or not, refining that >> latter threshold hasn't been a priority for me. >> >> Bob Menschel -- Please avoid sending me Word or PowerPoint attachments. See http://www.fsf.org/philosophy/no-word-attachments.html
