Does rspamd learn spam/ham from moving emails to spam folder/inbox?
It does not. Our perspective is that such a learning task is expensive on resources and only sounds good to the ears, but really does nothing to facilitate an improved spam filter. The only systems that we’ve found to be adequate at learning from being fed examples has been closed source systems like Gmail’s spam filter. Open source software performs quite differently. SpamAssassin being the worst, as it’s bayesian learning only seems to increase false positives. While rspamd is a bit better, we’ve never witnessed a reasonable improvement from it’s learning over time.
Instead, we’ve found it best to target efforts based on human recognition of spam trends. We’re not after one-off spam emails, but overall larger trends that we can tackle with custom rules. If you’re interested in contributing, you can do so here:
https://git.mxroute.com/jarland
An example of the custom rules can be seen here, where these phrases are rejected from the Subject header: