I've searches low and high for answers to this problem, but I believe they answers out there don't have regular predictable keywords to find them.
SA 3.0.1
Redhat FC2
Yes, that's what it's supposed to do in SA 3.x.
SA 3.0.x does not store bayes tokens in plain-text, it stores the SHA1 hashes of them. Since the tokens are hashed, it's impossible to derive what words they represent, so sa-learn --dump prints the hashes.
Unfortunately this makes sa-learn --dump a lot less useful, unless you're correlating against hashes output from spamassassin -D.
However, it makes SA's token matching much faster (fixed-length binary number compare instead of a string compare).
It also has the side effect of increasing user privacy, and the drawback of making it tough to double-check your own bayes DB.
Originally I responded to some emails from the Devs while 3.0 whas still in development, and they agreed to have an option to allow you to force SA to use plain-text bayes DBs, but apparently that option slowed things down too much.
