[ILUG] user-friendly, system/MTA/MDA level, spam filtering

[Declan Moriarty]

On Wed, 2006-06-28 at 20:33 +0100, paul at clubi.ie wrote:
> On Mon, 26 Jun 2006, Declan Moriarty wrote:
> 
> > Razor catches 40% or more. No false positives - mebbe one in years.
> Razor is good yes:
> 
> $ spamprobe dump ^Hx-razor
> 0.8187189    2246   31750  0x00000588  Hx-razor-warning_spam
> 
> That's 93% for me (but I get a decent amount of spam. 1k+ per day). 
> Razor is very good for feeding into Bayes filters.
> 
> > Spamc/spamd.
> 
> > Basic rulesets too soft on porn, bad language. Leave the threshold at 5
> > but work with your false negatives and tune. I have a score adjusting
> > local.cf which gets you a lot of the way. Delete rules which don't hit
> > anything.
> 
> Ok, how do you do that? Score adjusting and finding out the history 
> of rule matches?

Tuning based on false positives and false negatives mainly. Store marked
mail and use grep -c occasionally to enlighten you on 
1. Rule families (Some rulesets e.g. 5 similar fraud rules) only ever
score with one or two
2. Rules doing nothing.
3. Rules hitting ham.
4. Rules doing well for you.

Adjust accordingly. Tolerate false positives for a week.
-- 
        With Best Regards,

        Declan Moriarty.