I haven't touched any of the rules. I regularly train the bayesian filter
with false negative messages. I'm using local tests only.
My statistics show that only 60-65% of the spam messages are correctly
tagged as spam. I would like to hear from another spamassassin users if
they get similar figures.
I think many users if not most users are using some extra rules and also
running at least some net tests. Using some of the rules from
www.rulesemporium.com can help quite a lot with the newrt image-only spams
that are becoming quite popular. Using some of the assorted uribl type of
net tests can also help a very large amount.
I think most SA users that have net tests enabled and some SARE (rules
emporium) rules are probably getting well over 90% spam catch rates with
very few if any false positives.
IKn addition to training Bayes with FN messages, you should regularly feed
it a little ham and a little spam, just to keep reminding it what each one
looks like. Training on FNs will improve the hits on these, but won't
necessarily improve how Bayes does in general on spam. Almost all of your
spam should be hitting Bayes_80 or better. If not, feed it some more
representative spam. Also, most of your ham should be very close to
bayes_00.
Loren
