Teaching computers to automatically create piano arrangements from a melody is hard because high-quality examples of matched melodies and arrangements are rare.
The POP909 dataset provides 909 pop songs with professional piano arrangements made by musicians. The arrangements are stored in MIDI format, a standard digital music file, and are precisely time-aligned to the songs' audio and vocal melody.
It also includes extra labels for tempo, beat, and chords. This gives generative models, which are computer programs that create new content, the exact data they need to learn how tunes and accompaniments fit together, making automated arrangement more practical.