procmail/spassassin training session

Setup:
procmail v 3.22
Spamassassin v 3.3.2
Debian linux (testing)

I'm working in a procmail sandbox trying to teach spamassassin to do a
better job of recognizing spam. $sandbox/.procmailrc calls spamc like so:

,----
| :0fw
| | /usr/bin/spamc
| 
| :0:
| * ^X-Spam-Status: Yes   
| spama_spam_.in
|
| :0 
| post_sa.in
`----

I'm running procmail like this:

cat mixedmail_3000m| formail -e -s procmail -m ${sandbox}/.procmailrc

I have piles of spam/ham mix, some 60,000 messages, heavily leaning to
spam, that I've accumulated. 

I want a little coaching on running a spamassassin learning session.

First, I'd like to know if it matters that I have autolearn disabled in
SA config file for the duration?  I want to manually feed SA the spam
and ham so assuming I'd want autolearn off.

Another thing I wonder:

I planned to feed 3 thousand messages to SA then pick the falsely
filed ham out by hand and feed it again in an 
`sa-learn --mbox --spam falseham' command.

Now if I run the same several thousand messages thru SA as incoming
mail for the 2nd time, will SA do a better job of separating the ham
and spam? Or do I need to use different unprocessed mail for the
second run?

How many messages would make an effective learning session?