lowering your spam threshold howto

[Jo Rhett <jr...@netconsonance.com>]

R Lists06 wrote:
> Can you share your specific thought and implementation processes on this re:
> possibly going from 3.8 to 3.2 and how and why etc please?
 > We for one am interested as we are trying to move in that direction too.

It's very simple.  Tag messages above your soft limit and put them in a 
different folder.  Check the folder periodically for false positives. 
Try to identify why they are FP.

Look carefully at all of your normal mail, and confirm where it normally 
scores.

Lower your score limit to the lowest limit possible without creating FPs.

Keep watching.  When your FPs are down to less than 1 a week, you are 
probably safe leaving it that way.

-- 
Jo Rhett
Network/Software Engineer
Net Consonance

RE: lowering your spam threshold howto

[R Lists06 <li...@abbacomm.net>]

> 
> Hm.  Your experience differs from mine.  I tried using bayes, spent
> hundreds of hours training bayes with lots of good mail from
> archives, and lots of bad mail, and never got better than .5% (point-
> five or .005) difference in spam detection.  So we stopped using it.
> 
> In comparison, we've jacked the AWL score range and this works great
> for us.
> 
> --
> Jo Rhett

I shouldn't have been so flippant about AWL. What I was trying to do was say
that AWL without a good plan of mgmt can work against you.... much like
anything I spose.

I didn't even know how AWL worked or that it was working when I did my first
SA install.  Ooopps. :-)

Can you please specifically describe what you mean by jacked?

Or better yet, is there a specific howto so to speak for this approach of
yours please??

:-)

Thank you Jo.

 - rh

--
Robert - Abba Communications
   Computer & Internet Services
 (509) 624-7159 - www.abbacomm.net

Re: lowering your spam threshold howto

[Jo Rhett <jr...@netconsonance.com>]

On Feb 21, 2007, at 11:42 PM, R Lists06 wrote:
> Do you use bayes_auto_learn ?
> I am sure you know I don't mean AWL baloney.  ;-)

Hm.  Your experience differs from mine.  I tried using bayes, spent  
hundreds of hours training bayes with lots of good mail from  
archives, and lots of bad mail, and never got better than .5% (point- 
five or .005) difference in spam detection.  So we stopped using it.

In comparison, we've jacked the AWL score range and this works great  
for us.

-- 
Jo Rhett
Net Consonance : consonant endings by net philanthropy, open source  
and other randomness

RE: lowering your spam threshold howto

[R Lists06 <li...@abbacomm.net>]

> 
> It's very simple.  Tag messages above your soft limit and put them in a
> different folder.  Check the folder periodically for false positives.
> Try to identify why they are FP.
> 
> Look carefully at all of your normal mail, and confirm where it normally
> scores.
> 
> Lower your score limit to the lowest limit possible without creating FPs.
> 
> Keep watching.  When your FPs are down to less than 1 a week, you are
> probably safe leaving it that way.
> 
> --
> Jo Rhett
> Network/Software Engineer
> Net Consonance

Ok, great

Do you use bayes_auto_learn ? 

I am sure you know I don't mean AWL baloney.  ;-)

If so, What do you use for these settings then in the local.cf

bayes_auto_learn_threshold_nonspam -0.1

bayes_auto_learn_threshold_spam x.x

or do you hand train it all and no auto anything?

 - rh

--
Robert - Abba Communications
   Computer & Internet Services
 (509) 624-7159 - www.abbacomm.net