How Accurate is SpamAssassin's Bayes implementation?
This paper, by Gordon Cormack and Thomas Lynam of the University of Waterloo, details tests performed to measure accuracy rates of various open-source spam filters.
It compares the various modes of using SpamAssassin's Bayes implementation, then compares SpamAssassin's Bayes against other popular open source pure-learning spam filters (Bogofilter, Spambayes, SpamProbe, DSPAM, and CRM-114). SpamAssassin consistently does well, and is among the lowest in false positives (as of the July 1 2004 revision).
See page 12, 'Classification Performance - Pure Learning Filters', in particular.