On Mon, 29 Sep 2003, Jack Gostl wrote: > That new Bayes algorithm is mighty touchy. So far its tagged four real > messages with a BAYES_99, three of them today alone. In just five days > it has had twice the false positives that 2.55 had in four months.
I speculate that if your Bayes database was converted from 2.55, it'll be a little inaccurate until some training with 2.60 has occurred. I've had one false positive as a result of BAYES_99 -- which _is_ more than 2.55, as I have _never_ yet had a false positive with 2.55 -- but 2.60 autolearned that false positive as ham, and upon feeding it back through "spamc -R" to check, it was now BAYES_01. So I suggest teaching your false positives to 2.60, and it'll work things out pretty quickly. On the other hand, I've had a couple of true positives on BAYES_99 that were _also_ (incorrectly) autolearned as ham. In both the true and false positive cases this autolearning was because the message didn't hit any other rules. I'd rather retrain the false cases by hand and have the true cases ignored. So I've now got this in my .procmailrc: :0 ci * ^X-Spam-Status: Yes.*autolearn=ham | sa-learn --forget Back on the first hand again, though, I haven't had another mislearned true positive since adding that (that is, that rule has never actually fired). Which may be more representative of what spam I get than of improving Bayes accuracy, but ... ------------------------------------------------------- This sf.net email is sponsored by:ThinkGeek Welcome to geek heaven. http://thinkgeek.com/sf _______________________________________________ Spamassassin-talk mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/spamassassin-talk
