Dave Funk asked:
>Do you also exclude the Bayes score in your procmail recipe when
>you decide what to learn on?
>
>The Bayes auto-learn code explictly excludes the Bayes score when making
>its decsions to prevent positive feedback cycles.
After excluding certain senders (including this list), I automatically
sa-learn all of the leftover mail as either spam or ham according to
its final classification. My Bayes database seems to be fairly well
trained, and errors are unusual. I check my spambox daily for the
rare false positives, and I relearn those and the more frequent false
negatives as needed to squelch unwanted feedback. It works well for me.
>Why not do auto-learining, but learn to the journal file?
>As that's just an append it should be safe to do.
With just ordinary UNIX permissions available I have concluded that it
is unworkable to give both the spamd user and myself write access to
the Bayes database and journal file without also making them
world-readable and world-writable, which would expose the Bayes tokens
of my mail to all of the other users here. I need better privacy and
integrity than that.
>If it's a matter of write access over a network file system
>(EG NFS or CIFS) then you could hack up local journal files and then
>have a "sa-learn" cron job that collects the journal files and
>processes them.
Now you're talking! There might be a way to have spamd write to a
different journal file than is used when I run spamassassin. Using a
cron job is more complication than I want because this must scale up
to more than a thousand users, but some kind of support for multiple
journal files (one owned by me, the other by spamd) may be the ideal
solution in the long run.
:: Jeff Makey
[EMAIL PROTECTED]
