Pierrec and bane: My scripts are just a hodgepodge of bash and python. Happy to share (email me). I do not care much for MIDI piano; and since this music is typically monophonic, why not just use all the typical instruments? I generate the bodhran part from the MIDI and randomly choose to play a note, which kind of note to play, and whether to double up a note. I also give to each player the option to be late or early. Sometimes it gets a bit much, but is fun nonetheless.
bane: One of the reasons why I got into this music generation work is that it provides a sanity check of the internal models, just like speech recognition: look at the transcription to confirm the model is paying attention to relevant aspects. You may be interested in my "horse" article: "A Simple Method to Determine if a Music Information Retrieval System is a “Horse”" (http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6...). Another reason is that I like to compose, I like traditional music, and here we go!
Yenrabbit: Thank you! I agree completely. It is quite hopeful to believe a single artificial system will produce "music." It is merely shifting characters around according to a probabilistic model in light of constraints it has learned (such as four whole notes to a measure in common time), and it is up to musicians to "realise" it. Certainly, a lot is missing from the reduction to ABC. Music is much more than a sequence of arbitrary symbols. :)