You have two options here: learn such mail as spam => that's easy, create a catch-all mailbox on your destination mailserver and fetch all the mails from this mailbox and learn them as spam (e.g. see my script at my advancing PMG thread). However, you would also learn mails from people, who misspell mailboxes, so maybe look in your tracking center for "typical" spam boxes and set them up to learn. However, I already thought about that, but I'm afraid, that learning such mails as spam would have a big drawback: It's "dirty" spam and as/if this spam usually does not reach your valid mailboxes, your bayes filter will get worser by tagging just this "stupid shit" as spam and not real spam getting directly to existing mailboxes. I also started in the beginning by downloading spam archives to learn my spam filter, but it wasn't "my spam" it was "their spam" and so the bayes filter was really bad.
Another option you could choose and which would be much greater => setup a blacklist on such spam. I'm unsure, if it's good to set up on the sender IP as maybe you will add services like Google, Yahoo, etc. but maybe use the sender address (as long as it hasn't been spoofed this could help out). Maybe you want also to contribute your list.