Here's how we set up our site-wide SpamAssassin bayes database, including the ability for users to train it.
Create a new user spamd with home /home/spamd.
Perform the following as spamd:
cd /home/spamd
maildirmake -S Maildir, to create the shared SpamAssassin mailbox.
maildirmake -f SiteSpam -s write Maildir, to create a writable folder for misclassified spam.
maildirmake -f SiteHam -s write Maildir, to create a writable folder for misclassified ham.
Add the following to spamd's crontab to learn from and delete messages in those shared folders every five minutes:
0,5,10,15,20,25,30,35,40,45,50,55 * * * * sa-learn --spam --dir /home/spamd/Maildir/.SiteSpam/cur >/dev/null; cd /home/spamd/Maildir/.SiteSpam/cur ; ls | xargs -r rm 0,5,10,15,20,25,30,35,40,45,50,55 * * * * sa-learn --ham --dir /home/spamd/Maildir/.SiteHam/cur >/dev/null; cd /home/spamd/Maildir/.SiteHam/cur ; ls | xargs -r rm
The funny xargs-based way of deleting files is important, because ls has a hard limit on how many files it can handle, which we learned the hard way! Also, be sure there's a newline after the last line, or it won't be processed.
Modify /etc/spamassassin/local.cf with the directive:
bayes_path /home/spamd/
Modify /etc/default/spamassassin by adding -u spamd to OPTIONS.