SpamAssassin bayes learning works ?

Question

In the thread http://discussions.apple.com/thread.jspa?messageID=2343222&#2343222 I read about bayes autolearning with the two accounts "junkmail" and "notjunkmail".
However I thought that SpamAssassin automatically feeds the bayes filter automatically depending upon the score of each message.
But this thread makes me more in believing that without these two mail accounts SpamAssassin never learns.
So maybe someone can tell what is correct. And if it works only with these two special mail accounts, how can I achieve that SpamAssassin learns automatically (e.g. mails with score <= 1 are considered as ham, mails with score >1 is learned as spam) ? I have configured this similar on my old server, but have no idea how I made this years ago (maybe it was standard at SpamAssassin and this implementation is a special one by Apple).

Answer 1

You will need to feed it at least 200 ham and spam? After that, although I
This has been done; however I hadn't 200 ham messages, but learned around 1500 spams.
advise againt it, you can stop training it. However if you stop training it and
do not use rules, results are going to be poor.
Asking explicitly: Without junkmail/nonjunkmail mail accounts - does SA learn automatically ?
Are the tag levels set correctly in /etc/amavisd.conf?
Probably, I left them to the defaults. Here are the values:
satag_leveldeflt = -999
satag2_leveldeflt = 5.0
sakill_leveldeflt = satag2_leveldeflt (so also 5.0)
However... on my old server I have a mail tag like this one:
Yes, score=34.1 required=5.0 tests=BAYES_99, DATEIN_FUTURE_96_XX,FORGED_RCVD_HELO,FUZZY_AMBIEN,FUZZYVLIUM, FUZZYXPILL,HTML_50_60,HTML_FONT_LOW_CONTRAST,HTMLMESSAGE, HTMLTITLE_EMPTY,MIME_HTML_ONLY,RCVD_IN_BL_SPAMCOP_NET,RCVD_INXBL, SUBJALL_CAPS,URIBL_AB_SURBL,URIBL_JP_SURBL,URIBL_OB_SURBL,URIBLSBL, URIBLSCSURBL autolearn=spam version=3.1.3
Notice the "autolearn=spam" entry... this is what I miss at my new server as a proof that it has learned.