We use the lakh clean dataset https://colinraffel.com/projects/lmd/#get
and piano music: Combined from: https://github.com/bytedance/GiantMIDI-Piano and http://www.piano-midi.de/midi_files.htm
- Convert MIDI into .txt: ☑️DONE (only kept first X events, done in MidiConvert.py)
- Create a dataset: ☑️DONE (ConsolidateData.py)
- 2.1 i.e instruments used☑️, artist name☑️, song name☑️, (possible guessed genre from the previous info, give to gpt3.5 and let it guess, Not implemented)
- 2.2 Format this into a chat format, check out Resources.md. something like: ☑️"Create a song that " + {metadata} + "Response: Sure, here is Midi the code: {MIDI text}" for every song sample.
- Convert .txt into tokens with tokenizer?: 🟨Not started