What I'm worried about is that I have so many more SPAM than HAM messages. Is this dangerous?
No, in fact it's closer to optimal than a 50-50 mix is...
Remember, Bayes is a statistical system.. Statistics work best when they are as close to reality as possible. If most of your mail is spam, so should most of your training.
(You're currently at a 79% spam training ratio, if your real spam level is a bit over 70%, you're quite close to reality. That's VERY good.)
However, if you're not particularly close, don't worry about it, bayes is quite tolerant of reasonably wide imbalances.
In my case my learning is 96.4% spam training. My real-world spam percentage is 66%. Quite a difference there, but no problems.
