The quality of Bayes auto-learning depends a lot on the rules and thresholds 
you use.  On my gateway box with a lot of local rules as well as some add-on 
rulesets, I use auto-learn for the bulk of training, and hand-feed it only 
stuff that scores near the spam threshold.

After the last upgrade we had working Bayes within 12 hours on 10,000 msgs/day 
volume, and now it's really sharp with minimal hand feeding.  I just looked at 
an email that scored BAYES_99 and it really surprised me:

http://frodo.bruderhof.com/bayes99.txt

I would say there is no such thing as "Bayes poison".  This email was stuffed 
with innocuous text but it got zapped.  And I hardly ever see a Bayes-related 
false positive.

Pierre Thomson
BIC



-----Original Message-----
From: Matt Kettler [mailto:[EMAIL PROTECTED]
Sent: Tuesday, May 04, 2004 12:58 PM
To: Christopher Jett; [EMAIL PROTECTED]
Subject: Re: Newbie Question


At 11:16 AM 5/4/04 -0500, Christopher Jett wrote:

>If I do not manually train the bayes sytem on SA 2.63, how long does it take
>for the auto learn feature to start kicking in and bayes scores to start
>showing up?  I am running SA on a server-wide basis hosting multiple
>domains, so manually training is somewhat problematic at the moment.

I'd advise NOT using SA's bayes system on autolearn alone.. the quality of 
the bayes database produced isn't very good and generally ends up heavily 
poisoned.

Autolearning is a good supplement to reduce the frequency of training, but 
it's not really intended to be a sole source of training.


Reply via email to