> Your ratio of ham to spam shows you have a lot more ham than spam trained, > are you sure its not been learning spam as ham, so poisening your bayes > database.
I can't say I've looked at very many of the 100,000 hams. I have a quarantine area where I can skim through the spam and borderline stuff, but I don't keep a copy of the ham. However, to be learned as ham, the Nigerian messages would have to score below 0.5, and I don't think that's likely. Of course, there could be other messages that have some of the same tokens as Nigerian messages and that are being scored as ham. But they might actually BE ham. > Lower your BAYES_00 score? (Towards zero, that is) That's what I'm doing unless I can find something better.