> I have similar experience, and additionally bogofilter data files > (BerkeleyDB) are just *huge*. I've trained it with a corpus of > 25000 spam messages and 20000 non-spam messages, then I gave it a > test run on 200 previously unseen messages. 63% success rate, far > too low, and the databases exceeded 10M. SpamAssassin had 99.5% > percent success rate, but it is ridiculously slow compared to > Bayesian filters.
I have near-perfect bogofilter results with only 2426 non-spam messages and 3567 spam messages registered. The Robinson algorithm (-r) seems to be more effective than the default. For me, bogofilter is more accurate than spamassassin, thus I've dropped SA altogether for my personal use. The "foreign language problem" should no longer be an issue for anyone but CJK users.
