To answer my own questions:
It seems that by default SpamAssasin requires a minimum of 200 SPAM and
200 HAM messages in Bayes training database before it starts using Bayes.
This can be changed via SpamAssasins's local.cf by setting:
bayes_min_ham_num 50
bayes_min_spam_num 50
Also, I was able to copy Bayes databases from the old server to the new
one, but that revealed another issue with Bayes that was present on the
old server already. Databases were extremely 'imbalanced' - there were
millions of 'HAM' messages in the database and 'only' several thousand
of 'SPAM' messages - which is quite bad, as most tokens have been seen
in HAM messages much more often than in SPAM (or at least that's how I
understand the problem).
The cause for that was that I had Bayes auto learning enabled
while using Amavis as a centralized filter for all users and multiple
domains.
So for now I turned auto learning off - again in local.cf by setting:
bayes_auto_learn 0
I removed old Bayes databases to start over and am now training both HAM
and SPAM manually...
Regards,
Danilo
On 09.10.25 09:46, Danilo Godec wrote:
Hello,
I recently upgraded my central Amavis server from CentOS 7 to Rocky
Linux 9, which brought a few version changes in the relevant packages:
* amavis 2.12.3 -> 2.13.1
* spamassassin 3.4.0 -> 3.4.6
* rpsamd 3.8.4 -> 3.13.2
* dspam is no longer available on Rocky Linux (was 3.10.2 on CentOS 7)
* clamav 0.103.11 -> 1.4.3
Things mostly work, but I noticed that the new server never lists
'BAYES' in the tests list.
tests=[DKIM_SIGNED=0.1, DKIM_VALID=-0.1, HEADER_FROM_DIFFERENT_DOMAINS=0.001,
HTML_MESSAGE=0.001, HTML_MIME_NO_HTML_TAG=0.635, MIME_HTML_ONLY=0.1,
RCVD_IN_MSPIKE_H2=0.001, RCVD_IN_VALIDITY_RPBL_BLOCKED=0.001,
RCVD_IN_VALIDITY_SAFE_BLOCKED=0.001, SPF_HELO_PASS=-0.001, SPF_PASS=-0.001,
R_UNDISC_RCPT=3, R_DKIM_REJECT=0, DKIM_TRACE=0, MIME_HTML_ONLY=0.2,
ARC_ALLOW=-1, RCVD_TLS_LAST=0, ASN=0, FORGED_RECIPIENTS=2,
DMARC_POLICY_SOFTFAIL=0.1, FORGED_SENDER=0.3, DKIM_MIXED=0,
RWL_MAILSPIKE_POSSIBLE=0, TO_DN_ALL=0, R_SPF_ALLOW=-0.2, FROM_HAS_DN=0,
MIME_TRACE=0, RCVD_COUNT_FIVE=0, MISSING_XM_UA=0, TAGGED_FROM=0,
RCPT_COUNT_ONE=0, MID_RHS_NOT_FQDN=0.5, FROM_NEQ_ENVFROM=0, R_DKIM_ALLOW=-0.2]
I only trained the new server for a couple of days now, could it be
that it just didn't gather enough training data yet?
Best regards,
Danilo
