Thank you sir.
I agree that RagtimeBot is the best, and its dataset is indeed simpler in the sense that the rhythms and note structures are very similar among all data points. I.e. The left hand usually plays low bass note on beats 1 and 3, and higher chord on beats 2 and 4. The right hand is usually monophonic, with a syncopated rhythm.
The Beethoven dataset had the largest "note vocabulary" and thus had a harder time generating nice sounding music.
