On 17-06-15 20:00, Dieter Scholz wrote:
> Hello,
>
>>> My problem is: The bayes filter does (auto-)learn ham mails but no
>>> spam mails. In my logs I found spam mails that have a very high score
>>> and should be autolearned. I think my bayes setup is correct, because
>>> ham mails are learned as expected.
>> Autolearning is based on a different score to the one used for
>> classification. It uses the rule scores that would be used if Bayes
>> were disabled and ignores some type of rule altogether. For spam you
>> need 3 points from *both* the headers and the body.
>>
>> It probably is working correctly, as far as I can tell you haven't
>> actually established that no spam is being autolearned.
>
> Ok, there's a different score - I understand. But what worries me is,
> that I found mails in the log with a really high spam score and lots of
> rule matches. For all the (SA) marked spam mails I do not get a line in
> the logs with 'autolearn=no' as is the case for all ham mails. My mail
> server is now running for a week with hundreds of ham mails in the
> filter but no spams. It's a company server with 300 users and a long
> established MX record. So for me it is hard to imagine that there is no
> spam mail suitable for filter learning.
>
> I suspect the milter style of integration reject leads to a situation in
> which the spam learning part is suppressed. So there's no message about
> the mail being learned or not in the logs.
>
> What do you think? Any chance to debug this?
>
> Dieter
As far as I can see, you don't really know whether your setup is
actually learning spam or not. The logs just don't mention anything
useful about spam email. Could you check the bayes stats so we actually
know if there is a problem?
Please show us output of "sa-learn --dump magic", probably ran as the
user amavisd user.
Regards,
Tom
