Like a Bayesian learning system, SpamAssassin's GeneticAlgorithm requires a corpus of hand-classified mail. Our guidelines are (quoting and expanding on "masses/CORPUS_POLICY"):

• hand-verified as "spam" and "ham" (non-spam) piles – *not* just classified using existing spam-classification algorithms (such as SpamAssassin itself). Note that it's fine to use SpamAssassin to pre-filter them into the right piles, just make sure you scan the results "by hand" afterwards to verify that SpamAssassin made the correct diagnosis in each case.
  
• eliminate duplicates – there should be one and only one copy of any single email, whether spam or ham. (JustinMason: in my opinion, this isn't a hard and fast rule, as it can be very time-consuming. I'd suggest just removing dups where they all arrive at the same time, in sequence.)
  
• containing a representative mix of ham mail – that includes commercial-sounding-but-not-spam messages, legitimate business discussions (which may include talk of "sales", "marketing", "offers", bankruptcies, mortgages, etc), or verified opt-in mail newsletters. This is a very important point! Your ham corpus should contain as much ham as is possible, as close to ALL valid emails received by everybody as is possible, with only the exceptions noted here. ("as is possible" recognizes that for privacy and confidentiality reasons some ham cannot be stored anywhere but its destination email folder.)
  
• containing no old spam mail. Older spam uses different tricks and terminology, which will impact SpamAssassin's accuracy when it's filtering "live", new mail. Please try not to scan spam older than 6 months. For this purpose it may be useful to categorize your spam by month, and to regularly delete those files with the older spam.
  
• cleaned of viruses, and forwarded spam messages. These will skew the results.
  
• and finally, cleaned of discussion of spam or virus messages or signatures (such as SpamAssassin-talk or bugtraq mailing list messages). Even though they are ham, these often contain snippets of code that incorrectly trigger tests, and again will skew the results. (Rewriting the tests to avoid triggering on SpamAssassin-talk messages is not realistic!)
  

Once you run "mass-check" on a corpus (MassCheck), see the instructions in "CORPUS_SUBMIT" for details of how to verify that the top scorers are not accidental spam that got through.

(Aside: yes, it's "corpora". See PluralOfCorpus)