On Tuesday 04 October 2005 11:43 pm, Shane Chrisp wrote:
> > I've been using simscan to block messages that score 8 or higher for at
> > least 3 months now, and haven't had a single complaint.  The result is
> > we're blocking 57% of the inbound email, which greatly reduces the load
> > all around (fewer calls to vdelivermail, less load from users
> > downloading the spam, etc.)
> >
> > Based on my logs, we could block 50% by dropping at 12+.  I can't
> > believe that there would be a legitimate message scoring a 12+ that I
> > would not want to miss.

I believe you have these backwards.. dropping at 8+ includes all of the 12+ 
range :)

> Tom and Others,
>  How did/do you go about training SA? Just interested to hear what others
> are doing. We have been blocking at a score of 6.5 and do get the odd
> False Positive, though our customers dont seem to mind.

the key to training a bayesian classifier is using as much legitimate mail as 
possible.  The more spam you use (especially with spammers making up garbage 
filler paragraphs) the more you'll flag mail as spam.

I, personally, just round up all of my mailing lists and sa-learn --ham off 
that.  about 200k hams.  I haven't had a false negative in a long time, and 
never once have I had a false positive.


Jeremy Kitchen ++ [EMAIL PROTECTED]

In the beginning was The Word and The Word was Content-type: text/plain
  -- The Word of Bob.

Attachment: pgpHnPM5ITMtg.pgp
Description: PGP signature

Reply via email to