Re: [Evolution] Problem with junk and spam filters

[Brian Whiting <bwhiting cfl rr com>]

On Sun, 2005-12-04 at 01:38 +0100, guenther wrote:
> > > Slight correction: SpamAssassins built-in /Bayes/ filters need to be
> > > trained. The default SA rules work from the very first message, as do
> > > the network tests. Bayes needs to learn 200 Spam and Ham /each/, before
> > > kicking in (default install).
<snip>

> [1] Actually, "mails" are never learned. Their contents, the words are
> learned and identified with the overall score of the message. So
> according to the Bayes filter, some words are strong signs of being
> either Ham or Spam, whereas others aren't. (Basically, just keep in mind
> that words are learned, rather than entire mails.)

I find that the "learning" is still producing too many negatives.  I'm
getting repeat spam that is similar to "learned" messages.
I found my SpamAssassin settings in webmin & the "Hits above which a
message is considered spam" setting is at 5.  I don't know what a good
SA setting is for this.  I'd like to make it more discriminating in
reasonable steps, but need some practical pointers for a good way to
fine tune it.  Is there a generally accepted start setting for this? Are
the other settings more applicable to it?  I have:
Hits above which a message is considered spam 5
Whitelist score factor .5
Number of times to check From: address MX 2
Seconds to wait between MX checks 2
Skip RBL open-relay check? No
Seconds to wait for RBL queries 30
Number of Received: headers to check with RBL  2