Axb <axb.li...@gmail.com> writes: > On 11/12/2017 05:35 PM, micah wrote: >> David Jones <djo...@ena.com> writes: >> >>>> I am interested in seeing the bayes info in the database, because it was >>>> created years ago >>>> >>> >>> Spam changes all of the time so I train mine daily and manually expire >>> mine after about a month. Depending on your recipients, number of >>> mailboxes, and mail flow, you may be fine with not training that often >>> but I don't think tokens from years ago are going to be very accurate on >>> current mail flow. >>> >>> A large list of whitelist_auth entries with well-trained Bayes and you >>> can bump up the BAYES_* scores with nice results. >> >> How do you deal with a large user base with bayesian databases? It seems >> like having a shared one just gets useless fast, but allocating an >> individual database to each user is quite a hassle as well. >> > > Servicing +40k users I use a central Redis based Bayes db in autolearn > mode. Works for me.
I'm dealing with about 3x the users, I don't use auto-learn mode but let them train mistakes, but the whole thing is total garbage now and has to be reset again. Now it thinks with 99.9% probability that pgp encrypted emails are spam.
