Attempt at Naive bayes (bag of words model), using WEKA

Thanks to some very helpful redditors yesterday, I was able to convert some emails into text, then into an arff file. The two directories used as classes were orginally "spam" and "notspam" each of which contained a tiny (11 samples each) set of text files.

My goal is to use this tiny set to take some baby steps in running and understanding the software. What I'd really like to do is run a naive bayes on this arff file, to produce a "bag of words" model and then try it on some test data.

So my questions are:

1. Firstly, is there a tutorial for this specifically, if so, that would save us all a lot of time. My googling has not produced results, merely tutorials on the usage of weka on specific, pre-built data sets, which is not at all what i'm looking for. In addition, they don't use the naive bayes either.

2. What are my next steps using this arff file? Surely it has to be filtered and I have to make sure the features are correctly selected. Is the arff file I've built going to be adequate for some very small short tests, just to see how it all works?

3. I noted when i was using textdirectoryloader that it specifies that UTF-8 is the way to go (and this is what I used to do the conversion, from the CLI), especially since my endgame will involve testing html or email files, with the header included, but the tags removed. Do I need more preparation, like removing the headers themselves, or will this be adequate?

Thanks in advance for your time.