That makes sense. We're still very much in the testing phase of SA so I
don't really have a good way for users to report false positives and
negatives. A few weeks ago I saved all spam to a folder, weeded out any
FPs, and learned it all as spam. For ham I grabbed my mailbox and a few
others. Mail's a pretty touchy subject so I don't know how receptive
people would be to us grabbing their legit mail to train Bayes. I'm doing
expiry runs daily. Should I consider re-learning my spam corpus that I've
saved up?
Don
On Mon, 3 May 2004, Matt Kettler wrote:
> At 02:46 PM 5/3/04 -0400, Don Newcomer wrote:
> >m having a problem that are a little similar to one reported earlier
> >today but not quite the same. After seeing a strange-looking false
> >positive, I realized that there were no Bayes scores that would have offset
> >the high score from the rulesets. I did a count on my MailScanner logs and
> >found that 1/3 of the messages labeled as spam had no Bayes scores. Bayes
> >is working fine otherwise. I'm running SpamAssassin 2.63 with MailScanner
> >4.29.7 (soon to be 4.30). Any idea what's happening here? Thanks in
> >advance.
>
> Lack of bayes score isn't abnormal. Normally it's a sign your bayes
> training is lacking.
>
> If there is nearly and exact 50/50 probability of spam ( between 49.99%
> and 50.01%), or if there are no token hits at all, no bayes score will be
> reported by SA.
>
> However it's a bit unusual to have 1/3 of your messages not hit bayes at
> all, unless your training is VERY deficient. How often do you train
> manually? What do the stats look like if you do a sa-learn --dump magic?
>
>
>
>
