[Rspamd-Users] Impact of Whitelisted Messages on Bayes/Neural Learning in rspamd

[Denny Friebe]

Hello Florian,
Thank you for the suggestion. This would certainly serve as a temporary 
workaround for the spamreports.

However, I frequently send various monitoring notifications through my 
mailserver, and I would like to completely exclude those messages from 
the learning process as well. Therefore, I am looking for an alternative 
approach.

While reviewing the documentation again, I came across an example under 
Configuration → User settings which uses a whitelist with the flag 
"want_spam = yes". According to the description, this should provide a 
complete bypass of rule evaluation.

Would I be correct in assuming that such a configuration would also 
prevent these messages from being included in the autolearning process?

As far as I understand, this method would also disable logging for those 
messages in the web interface. However, I would prefer to keep logging 
enabled, so I can retain a full overview of both incoming and outgoing 
messages.

Alternatively, this might also be achievable using the symbols_enabled, 
groups_enabled, or the symbols_disabled / groups_disabled arrays. But I 
am unsure, which specific symbols or groups are responsible for learning 
and logging functionality.

Am 2025-08-06 17:18, schrieb Florian Piekert:
> Hello Denny,
> 
> how about attaching them as a zip/7z archive file instead?
> 
> Especially due to this "fear" of tainting the neural learner I do this 
> (atm without zip/7z, but as a base64 encoded attachement (mpack)).
> 
> Florian
> 
> Am 06.08.2025 um 17:10 schrieb Denny Friebe via Users:
>> Hello everyone,
>> I’m currently using a cronjob to send out a daily spam report to 
>> mailboxes that do not use IMAP. This report is intended to inform 
>> users about newly received messages that were classified as spam. It 
>> includes details such as the reception time, Envelope-From, From, 
>> Subject, and the IP address of the sender, all formatted in a table.
>> 
>> Initially, these reports were themselves marked as spam by rspamd. To 
>> address this, I created a multimap based whitelist that includes the 
>> full sender email address:
>> 
>> WHITELIST_FROM {
>>      type = "from";
>>      map = "$CONFDIR/local.d/whitelist_from.map";
>>      description = "Local from whitelist";
>>      action = "accept";
>> }
>> 
>> This works as expected. Rspamd now accepts the reports without 
>> flagging them as spam.
>> 
>> However, I am wondering how this affects the Bayes or neural learning 
>> process. (i use autolearning)
>> Are these whitelisted messages still included in the classifier 
>> training despite being accepted via multimap?
>> If so, is there a recommended way to exclude them from training to 
>> avoid negatively impacting the learning accuracy?