I've been using spam-stat.el for ages, but was unhappy that it only
had a success rate of about 97%. Checking a bit this was almost
entirely because it did no decoding of mime (or base64) encoded
articles. I've modified it to decode mime (if you don't like this it
can be controlled by customizing the spam-treat-mime-function to
nil). 

After retraining, I now have a false-positive rate of less than .08%
(no false positives on my test directory of 1300 ham emails) and a
success rate of detecting spam of about 99.8%, which is as good or
better than any of the other Bayesian filters I've played with. 

Diffs attached.