rspamd
[Feature] Pre-trained/seeded Bayes/Neural
Summary
Support for pre-trained/seeded Bayes/Neural.
Ideally this would allow collecting a long-term training set of spam/ham that is then used to generate Bayes/Neural (so, based on known good data) and then continue using it as usual (with Bayes/Neural autolearn).
Motivation
It takes a bunch of fine-tuning to get a good Neural/Bayes set-up with autolearn that remains stable over time. If something goes wrong it can require a full reset, this can be non-optimal.
If the goal is to achieve long-term Bayes/Neural to catch certain phish/spam campaigns that happen a few times each month, then having to reset either can become very detrimental.
Describe alternatives you've considered
Using an existing Bayes/Neural for a longer time - might go haywire if either gets taught wrong (automatically). But disabling autolearn (or the likes) would reduce effectiveness significantly.
I think I just kind of asked somewhat the same question. So I think this is a very good idea obviously https://github.com/rspamd/rspamd/issues/4776
