On Fri, Oct 21, 2005 at 05:15:56PM +0200, Alexander Dalloz wrote: > Am Fr, den 21.10.2005 schrieb D. D. Brierton um 17:02: > > > I'm using FC4 with spamassassin-3.0.4-1.fc4. fetchmail delivers mail to > > a locally running postfix. spamd is running as a service, and spamc is > > called by procmail on my mail. My setup is almost identical to that > > desribed here: > > > > http://wiki.apache.org/spamassassin/UsedViaProcmail > > > > However, despite the fact that I have trained spamassassin on a vast > > amount of both ham and spam using sa-learn, I suspect that Bayesian > > testing is not being applied. I became suspicious that this might be the > > case after receiving over a dozen almost identical messages and despite > > training spamassassin on them they are still not being identified as > > spam. So I started looking at the headers that spamassassin adds to each > > message more closely. Here is the header it added to a recent message > > from this list: > > > > X-Spam-Status: No, score=0.0 required=5.0 tests=RCVD_BY_IP > > autolearn=failed version=3.0.4 > > You are irritated by "autolearn=failed"? Autolearning is not the same as > using the Bayesian features. Though the former requires the last to be > active. And autolearning does not use each mail checked, it only applies > to messages with more or less reliable spam / ham score values. > > > And here is an example of an incorrectly identified spam message: > > > > X-Spam-Status: No, score=2.8 required=5.0 tests=HELO_DYNAMIC_IPADDR, > > RCVD_BY_IP autolearn=no version=3.0.4 > > Did you set in local.cf something like following? > > use_bayes 1 > auto_learn 1 > bayes_learn_to_journal 1 > bayes_file_mode 0600 > bayes_path /var/spool/MIMEDefang/bayes > > > Best, Darren > The lines like the ones above appear in: .spamassassin/user_prefs assuming these are personal user settings. In my instructions auto_learn appears as bayes_auto_learn. The last three lines have default values so I would assume they are not needed. One more thing the option: bayes_auto_learn_threshold_spam needs to be considered since a spam score below this threshold is not auto-learned to be spam. Its default in 12.0. Similarly there is a:bayes_auto_learn_threshold_nonspam. Default is 0.1 -- ======================================================================= "Pascal is Pascal is Pascal is dog meat." -- M. Devine and P. Larson, Computer Science 340 ------------------------------------------- Aaron Konstam Computer Science Trinity University telephone: (210)-999-7484
