On Mon, 19 Apr 2004, Joe Conway wrote:Marc G. Fournier wrote:Huh? I just use Spamassassin myself, with Razor/Pyzor/DCC and Bayes all enabled ...
I use exactly the same setup. But recently I've noticed that the spammers are getting smarter -- I think 20% of it is slipping by the filters. I'm going to need something better.
do you force learn those spam that get through the cracks? I get about 20 or 30 messages that slip through the cracks, which I process through with sa-learn nightly ...
Sorry to drag this OT thread on even longer, but it seems to be a topic many are interested in ;-)
I wanted to report back that after just 2 days of forced (supervised) learning, the bayesian filter is now nailing about 99% of all spam. *Many, many, thanks* for the suggestion.
But I wonder why the autolearn feature is so conservative? At this point I'm getting lots of stuff like this:
X-Spam-Status: Yes, hits=5.8 required=2.5 tests=BAYES_99,HTML_FONT_BIG, HTML_MESSAGE autolearn=no version=2.63 X-Spam-Report: * 0.1 HTML_MESSAGE BODY: HTML included in message * 0.3 HTML_FONT_BIG BODY: HTML has a big font * 5.4 BAYES_99 BODY: Bayesian spam probability is 99 to 100% * [score: 1.0000]
Notice that, even though I get a hit on BAYES_99, I still get autolearn=no. Ah well, I guess I should be asking that question of the SpamAssassin guys. Also notice that this sucker would have gotten through with a score of only 0.4 had it not been for the bayesian filter.
Again, thanks.
Joe
---------------------------(end of broadcast)--------------------------- TIP 9: the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match
