I setup spamassassin with a site-wide bayes database. Users are reporting their 
own spam, and after being approved by an administrator, that spam is used to 
train the spamassassin bayes database.

Because I have users reporting spam into a global bayes database, I want the 
learner to ignore any e-mail addresses of my users in the learning, because if 
one user happens to report lots of spam, bayes would learn that their address 
means spam. I don't want this.

I have already excluded the To, Cc, Bcc headers using the base_ignore_header 
config, however e-mail addresses show up in my Received header like the 
following and can show up others places too.

Received: from w3.drh.net ([64.21.76.5])
          (envelope-sender <dharris@drh.net>)
          by secondary.scan1.myactv.net (qmail-ldap-1.03) with SMTP
          for <test.mss2@mail.myactv.net>; 20 Jul 2005 18:12:32 -0000

So, I created a patch that applies the below regular expression to any text 
before it tokenized by bayes to wipe out the username:

s/[a-z0-9][a-z0-9\_\.-]{1,48}\@
(myactv.net|mail.myactv.net|mss1.myactv.net)/MYACTVREPLACEDUSERNAME\@myactv.net/
gi;

Because I have multiple MX servers, I also used this regular expression to 
solve the problem described here http://wiki.apache.org/spamassassin/BayesBitMe

s/scan\d.myactv.net/scan1.myactv.net/g;

A configurable way rewrite text before tokenization would be appreciated.

Also note that crm114 (http://crm114.sourceforge.net/) has a feature to do this 
same thing.

Here is my patch to add this feature manually:
http://www.davideous.com/qmail/Mail-SpamAssassin-3.0.4-antietam-bayes-
customizations-040719-just-rewrite.patch