On Thu, 15 Jun 2006 13:17:51 -0700, "Harris, Jason \(DIS\)" <[EMAIL PROTECTED]> wrote:
>I'm wanting to know many times per year do SA admins have to retrain ? > >Our setup sends mail to SA client for a score, then depending on score >stores a backup of the mail in spam/ham mail folders for later review in >case a mistake is made. We train SA and it detects flawlessly at this >beginning time; the good mail numbers about 2000 per day and the spam >rates at about 10000 per day. In three months of time, SA is letting >most of the spam through, the rates I've listed above being reversed. > >We keep a month of mail around for retraining, which is lot of work to >go through. I was just wondering how ofter others have to do the same >thing. Thanks! > >SA Version 2.64 >skip_rbl_checks 0 >bayes_auto_learn_threshold_spam 7 >use_bayes 1 > > >Jason Harris My Bayes DB is over 2 years old, I've done no mass retraining apart from a few initial attempts. Since then I don't use auto_learn. When I did I had the score set at 30 or above. With a score as low as 7 you are likely to get all sorts of bayes poisoning spams trained in which may explain why you are needing to retrain so often. 3.13 is the current version, I'm currently running 3.11 & 3.12 with various SARE rulesets and I get very few FP's. Those that are found on manual checks are retrained as ham, any spam getting through to the user is retrained as spam. That combination has worked here and with at least 2 other colleagues for about the same length of time (2 years +) HTH Nigel
