I just discovered the antispam plugin (http://johannes.sipsolutions.net/Projects/dovecot-antispam) and set it up successfully. This looks like a really simple an efficient way of handling retraining.
For testing purposes I was moving a (correctly classified) mail from a user's INBOX to the SPAM folder where it was handed over to DSPAM for retraining. I verified the retraining (it has actually being processed 4 time due to the setting "TestConditionalTraining on"), but when verified manually the mail was still classified as "Innocent".
E.g.
# dspam --client --user myuser --mode=notrain --stdout --classify < mail_which_should_now_be_spam X-DSPAM-Result: myuser; result="Innocent"; class="Innocent"; probability=0.0000; confidence=0.51; signature=4991de6d180672059758908
Now, this is something which is happening from time to time and results in similar SPAM getting through several times until DSPAM correctly recognizes the SPAM (even though "TestConditionalTraining" is activated).
My question is the following: Is it possible to first retrain (with --mode=error) and afterwards corpus feeding the mail until it
a) is correctly classified b) has a (configurable) confidence of say 0.8
If a retrained mail gets through a second time the whole process of re-learning and training doesn't make sense in a user's mind and gives the impression that the SPAM setup doesn't quite work (even though the hit rate may be well over 98%).
Thanks, Reto