1) How can we train the spamassassin using bayesian to FILTER ALL
OUTGOING AS WELL AS INCOMING messages from my server.
That depends on your MTA. Some make it easy (Postfix), some make it
difficult (Sendmail), some you just shouldn't be using (Qmail). (That
last one was a joke, people.)
2) Some really Best Practices for implementing and running Spamassassin.
Like the other poster said, consult the archives. Also look at the wiki.
But generally...
1. Properly train your Bayesian filter.
2. Use the default 5 spam threshold.
3. Mark messages as spam only, and let the users be responsible for
their own deletion.
