While rspamd is objectively more efficient, we've reverted our DirectAdmin servers to use SpamAssassin. The most significant reason being that the SpamAssassin user configuration in the DA panel would not work fully as expected (ex. no wildcards on white/blacklist), and this led to us spending too much time trying to work with users to achieve their goals here (or to explain why they couldn't and were better off than if they were able to). The most wise decision appears to be to restore that page to work as users expect.
We expect a very functional configuration that helps reduce spam to look something like this: https://www.dropbox.com/s/5mmd017tfvvzw2b/Screen-Shot-2020-01-22-11-05-52.59.png?dl=0
If you see false positives to any excessive degree, maybe try "Medium Threshold" instead.
No you do not have the ability to train the filters, bayesian learning is not effective or efficient and we're not using it to train when you move email to/from the junk/spam folder.