BTW, Matt was right in his assumption below. AWL worked correctly on my test. I intentionally contrived 2 emails from the same fake address. The first was inoccent the 2nd was the same text plus few known spammy words and phrases.
On 8/24/05, Matt Kettler <[EMAIL PROTECTED]> wrote: > jdow wrote: > > Ilan, you could adopt my strategy and simply turn off auto-whitelist > > and delete the auto-whitelist file. I've seen too many mis-trained > > auto-whitelists mentioned on this list to be at all comfortable with > > it. The same can be said for auto-learn with Bayes. > > > > Are you sure they were mis-trained? > > Or were you just seeing cases of negative AWL scores in spam (which is > perfectly > normal)? > > http://wiki.apache.org/spamassassin/AwlWrongWay > > > > At the VERY least set the thresholds for auto this and that MUCH wider > > than they come stock. > > That won't affect the AWL at all. That only affects bayes. The AWL is an > averager, so it "learns" every message. > > > > > It appears you have a spam message yet auto-whitelist thinks it is > > ham. > > No. The second message isn't spam. It's a spammy ham, or at least so the Ilan > implies: > > "Anyway, I did as advised and ran spamassassin -D < test (instead of the > --lint option) and I ran it twice on 2 messages from the same address > (2nd was spammy). This way it does work as advertised" > > Note Ilan did not say it was spam, just spammy. Also note that Ilan considers > this part to be *correct* behavior. > > So that means his AWL *correctly* deducted points from a message that would > have > been a FP otherwise. > > It's possible Ilan is using intentionally contrived emails here to force the > case. > > (If it really was ham, you found a reason to sort spam into a > > spam mailbox and at least glance at the trash before tossing it.) > > -- Ilan Aisic Registered Linux User 8124 http://counter.li.org
