spamassassin: unreasonable scores for BAYES_00 and BAYES_05 ?

[hellfire]

In /usr/share/spamassassin/50_scores.cf are the following 2 lines:

score BAYES_00  0  0 -1.5   -1.9
score BAYES_05  0  0 -0.3   -0.5

Since I notice a some spam with image-only content(which is spam-rated for that reason), the bayes element of pmg adds high negative score and thus allows those mails to get through. I'm wondering if this is not a very reasonable default and this should be corrected?

(In my installation I reset the scores to 0, so it stays neutral and only negatives results matter).

 

[Stoiko Ivanov]

I think the issue here is that your bayes filter is trained wrong - I would suggest to simply disable bayes filtering (GUI->Configuration->Spam Detector->Options)

I hope this helps!

 

[hellfire]

Hmmm. I wonder if I did not understand the documentation right:

Page 2 PMG - Documentation

Bayesian Filter - Automatically trained statistical filters
Certain words have a higher probability of occurring in spam emails than in legitimate emails. By being
trained to recognize those words, the Bayesian filter checks every email and adjusts the probabilities
of it being a spam word or not in its database. This is done automatically.

I assume this works like that: When ever a spam message is detected by other means of the spam detection, the Bayesian Filter is trained so that the figured out bad message is used for training the bayesian filter automatically.

I did not do any manual training myself.

Is my understanding correctly here?

 

[Stoiko Ivanov]

I assume this works like that: When ever a spam message is detected by other means of the spam detection, the Bayesian Filter is trained that the figured out bad message is used for training the bayesian filter automatically.

The documentation needs an update - we're currently planning on disabling the autolearn feature of Bayes in spamassassin (which is something I'd rather do on a major version change to give it more visibilty) - mostly due to our experience with it in practice (from reports here and from our enterprise users) - the rationale is quite well summed up in this thread:
https://lists.apache.org/thread/0w4kp9vhmlqjp424oxf1qtn3t09c97bc

So - I would suggest to disable bayes and see how it works in your case (I would expect better than with it enabled)

I hope this explains it!