I have read several notes (and the web page) indicating that one should strive for 1:1 ham / spam in training.

Approximately, yes. 1::1 is ideal, but you should be fine with 2::1, 1::3, etc. People with ratios like 65::1 or 431::1 tend to get weird results. The golden rule is that if things are working, then you shouldn't need to change them.

Given that I use the web interface for training, I am not clear on how I would go about doing that.

Indeed, it is somewhat of a difficult task. Training only on mistakes and unsures tends to help in my experience. Adjusting the thresholds can help, too - particularly the spam threshold, which can probably be reasonably safely dropped down to around 80% once the classifier is trained. If you get a lot of duplicate spam, then only training on one of the duplicates will help (assuming that the problem is too much spam). You could do training on fp, fn, unsures and a few ham every now and then to get the balance up. Is that any help?

Also, thank you very much for getting the 1.0.4 fix out. That is a big help.

You're welcome. =Tony.Meyer -- Please always include the list (spambayes@python.org) in your replies (reply-all), and please don't send me personal mail about SpamBayes. http://www.massey.ac.nz/~tameyer/writing/reply_all.html explains this.