asigalov61/Discover-Piano

Discover Piano

Ultimate pre-tokenized solo Piano MIDI dataset for symbolic music AI and MIR purposes

---

Installation and use

---

Load dataset

#===================================================================

from datasets import load_dataset

#===================================================================

discover_piano = load_dataset('asigalov61/Discover-Piano')

dataset_split = 'train'
dataset_entry_index = 0

dataset_entry = discover_piano[dataset_split][dataset_entry_index]

midi_hash = dataset_entry['md5']
midi_score = dataset_entry['score']

print(midi_hash)
print(midi_score[:15])

---

Decode score to MIDI

#===================================================================
# !git clone --depth 1 https://github.com/asigalov61/tegridy-tools
#===================================================================

import TMIDIX

#===================================================================

def decode_to_ms_MIDI_score(midi_score):

    score = []

    time = 0
    
    for t in midi_score:

        if 0 <= t < 128:
            time += t * 32
    
        elif 128 < t < 256:
            ptc = (t-128) 
    
        elif 256 < t < 384:
            dur = (t-256) * 32

        elif 384 < t < 512:
            vel = (t-384)

            score.append(['note', time, dur, 0, ptc, vel, 0])

        elif 512 <= m < 845:
            chord_tok = m-512
            chord = TMIDIX.ALL_CHORDS_SORTED[chord_tok-12] if chord_tok > 11 else [chord_tok]

        elif 845 <= m < 973:
            bar_tok = m-845

    return score
    
#===================================================================

ms_MIDI_score = decode_to_ms_MIDI_score(midi_score)

#===================================================================

detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(ms_MIDI_score,
                                                          output_signature = midi_hash,
                                                          output_file_name = midi_hash,
                                                          track_name='Project Los Angeles'
                                                          )

---

Dataset pipeline

• Source MIDI dataset: Discover MIDI Dataset
• Dataset was created using the following pipeline code:

import os
import TMIDIX # v26.7.20

clean_midis = TMIDIX.read_jsonl('Discover-MIDI-Dataset/DATA/Files Lists/clean_midis_files_list.jsonl')

filez = [os.path.join('Discover-MIDI-Dataset/MIDIs', f['md5'][0], f['md5'][1], f['md5']+'.mid') for f in pool]

CLEAN_INSTRUMENTS = TMIDIX.LEAD_INSTRUMENTS+TMIDIX.BASE_INSTRUMENTS

def process(input_midi):

    try:
        raw_score = TMIDIX.midi2single_track_ms_score(input_midi)
        
        escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True, apply_sustain=True)[0]
        
        escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes, timings_divider=32)

        escore_notes = [e for e in escore_notes if e[6] < 80 and e[6] in CLEAN_INSTRUMENTS]

        escore_notes = TMIDIX.solo_piano_escore_notes(escore_notes)

        escore_notes = TMIDIX.remove_duplicate_pitches_from_escore_notes(escore_notes)

        escore_notes = TMIDIX.fix_escore_notes_durations(escore_notes, min_notes_gap=0)
        
        cscore = TMIDIX.chordify_score([1000, escore_notes])

        fixed_score = []

        for c in cscore:
            c.sort(key=lambda x: -x[4])

            tones_chord = sorted(set([p[4] % 12 for p in c]))

            if tones_chord not in TMIDIX.ALL_CHORDS_SORTED:
                tones_chord = TMIDIX.check_and_fix_tones_chord(tones_chord, use_full_chords=False)

            for e in c:
                if e[4] % 12 in tones_chord:
                    fixed_score.append(e)

        vels = [e[5] for e in fixed_score]
        avg_vel = sum(vels) / len(vels)

        if len(set(vels)) == 1:
            fixed_score = TMIDIX.humanize_velocities_in_escore_notes(fixed_score)

        if avg_vel < 80:
            TMIDIX.adjust_score_velocities(fixed_score, 100)

        cscore = TMIDIX.chordify_score([1000, fixed_score])

        score = []

        abs_time = 0
        pbar = -1

        pc = cscore[0]

        for c in cscore:
            c.sort(key=lambda x: -x[4])
            
            if abs_time // 128 > 127:
                break

            if abs_time // 128 > pbar:
                score.append(min(127, (abs_time // 128))+845)
                pbar = min(127, (abs_time // 128))

            tones_chord = sorted(set([p[4] % 12 for p in c]))

            if len(c) > 1:
                chord_tok = TMIDIX.ALL_CHORDS_SORTED.index(tones_chord)+12

            else:
                chord_tok = tones_chord[0]

            dtime = max(0, min(127, c[0][1]-pc[0][1]))
            score.append(dtime)

            score.append(chord_tok+512)
            
            abs_time += dtime

            for e in c:
                score.extend([max(1, min(127, e[4]))+128, max(1, min(127, e[2]))+256, e[5]+384])

            pc = c

        if len(score) > 255:
            return os.path.splitext(os.path.basename(input_midi))[0], score

    except:
        return None

---

Changelog

Version 26.7.24

• Dataset now features the best possible encoding scheme which should greately improve the results.
• This is a breaking change from the original version so if you need the original version, please pull it from the respective commits/DOIs.

---

Citations

@misc{DiscoverPiano2026,
  title        = {Discover Piano: Ultimate pre-tokenized solo Piano MIDI dataset for symbolic music AI and MIR purposes},
  author       = {Alex Lev},
  publisher    = {Project Los Angeles / Tegridy Code},
  year         = {2026},
  url          = {https://huggingface.co/datasets/asigalov61/Discover-Piano}

@misc{project_los_angeles_2025,
    author       = { Project Los Angeles },
    title        = { Discover-MIDI-Dataset },
    year         = 2025,
    url          = { https://huggingface.co/datasets/projectlosangeles/Discover-MIDI-Dataset },
    publisher    = { Hugging Face }
}

---

Project Los Angeles

Tegridy Code 2026