Galaxy MIDI Dataset

The largest and most comprehensive MIDI dataset in the known galaxy

---

---

A lifetime of music: 19,893,571 unique, standardized MIDI files with pre-computed features, midisimx embeddings, lyrics, karaoke data, and genre labels, ready for Music AI and MIR.

---

Contents

• Introduction
• Abstract
• Overview
• Key features
• Installation and use
• Documentation
• Dataset structure
• Dataset statistics
• Dataset integrity verification info
• License
• Citations

---

Introduction

The Galaxy MIDI Dataset is the flagship symbolic music corpus from Project Los Angeles / Tegridy Code — the direct continuation of a line of record-setting open datasets that includes the Tegridy MIDI Dataset and the Discover MIDI Dataset (see Citations).

At its core are 19,893,571 unique, standardized, and original MIDI files — but Galaxy goes far beyond raw MIDI. Every portion of the dataset is enriched with pre-computed data: comprehensive MIDI features, embeddings, file identifications, genre/artist/song labels, karaoke tracks, aligned English lyrics, curated subsets, and high-quality soundfonts for rendering.

Working with a dataset of this scale is made effortless by the companion galaxymidi Python package: downloading and extracting the entire 152GB corpus takes just a few lines of code. The embeddings portion of the dataset is powered by midisimx, which galaxymidi depends on and installs automatically.

The entire dataset is released under the CC BY-NC-SA 4.0 license, making it free for non-commercial research, education, and creative use with attribution.

---

Abstract

We present the Galaxy MIDI Dataset — to our knowledge, the largest and most comprehensive symbolic music dataset ever released. The dataset consists of 19,893,571 unique, standardized, and original MIDI files (152 GB), accompanied by rich pre-computed annotations and derived data: 19,893,571 comprehensive MIDI feature sets, 15,591,559 embeddings computed with the midisimx embedding model, 9,361,365 file identifications, 89,569 files with genre, artist, and song labels, 154,990 karaoke MIDIs, and 47,069 files with corresponding English lyrics. Curated subsets include 3,667,206 MIDIs with monophonic melodies, 1,355,292 solo melody MIDIs, and 742,915 solo drum-track MIDIs. The dataset ships with eight curated JSONL file lists for memory-efficient subsetting, four production-grade soundfonts for rendering and auditioning, and full integrity verification via MD5 and SHA256 checksums. The entire corpus is released under CC BY-NC-SA 4.0 together with a lightweight companion Python package (galaxymidi) that handles downloading and fast parallel extraction in a few lines of code. The Galaxy MIDI Dataset is designed to support music information retrieval (MIR), symbolic music generation, music understanding, representation learning, and large-scale pretraining for Music AI.

---

Overview

The dataset is organized into a small number of self-describing top-level directories:

Directory / file	Description	Scale
MIDIs	The complete corpus of unique, standardized, original MIDI files	19,893,571 files
Features	Pre-computed comprehensive MIDI features for all MIDIs in the dataset, merged into a single container	19,893,571 entries
Embeddings	Pre-computed midisimx embeddings for the clean MIDI subset	15,591,559 vectors
Files Lists	Curated JSONL file lists for fast subsetting and streaming	8 lists
Identified	MIDI file identifications	9,361,365 entries
Karaoke	Karaoke MIDI data	154,990 MIDIs
Genres	Genre, artist, and song labels	89,569 MIDIs
Lyrics	Aligned English lyrics	47,069 MIDIs
SoundFonts	High-quality SF2 soundfonts for rendering and auditioning	4 soundfonts
Artwork	Official dataset artwork	—
Code	Bundled source code: midisimx, tegridy-tools, and license	—
ATTRIBUTION.txt	Source attribution information	—

The curated Files Lists provide instant access to all major subsets without scanning the full corpus: all, clean, aligned, drum_track, genres, karaoke, mono_melodies, and solo_melodies.

Tooling

• galaxymidi — companion pip package for one-command download, fast parallel extraction, and data loading. Installable via pip install galaxymidi.
• midisimx — MIDI similarity and embedding model used to compute the pre-computed embeddings portion of the dataset. Installed automatically as a dependency of galaxymidi.

---

Key features

• 🌌 Galaxy scale — 19,893,571 unique MIDI files / 152GB compressed: the largest MIDI dataset of its kind
• 🎼 Unique, standardized, and original — fully deduplicated and standardized corpus
• 🎛️ 19,893,571 pre-computed comprehensive MIDI features for all MIDIs in the dataset — no extraction pipeline required
• 🧬 15,591,559 pre-computed midisimx embeddings — ready for similarity search, retrieval, and downstream modeling
• 🔎 9,361,365 identified MIDIs — matched and identified files
• 🎤 154,990 karaoke MIDIs and 47,069 MIDIs with aligned English lyrics
• 🏷️ 89,569 MIDIs with genre, artist, and song labels
• 🎹 Rich curated subsets — 3,667,206 mono-melody MIDIs, 1,355,292 solo melodies, 742,915 solo drum tracks
• 📋 8 curated JSONL file lists — instant, memory-efficient subsetting
• 🔊 4 production-grade soundfonts for immediate rendering and auditioning
• 🧰 Companion pip package — download and extract the whole dataset in two lines of code
• ✅ Full integrity verification — MD5 and SHA256 checksums included
• ⚖️ Clearly licensed — CC BY-NC-SA 4.0

---

Installation and use

Requirements: Python 3.x with pip. Please note that the dataset size is 154GB compressed and 330GB uncompressed, so make sure you have sufficient disk space (~500GB) and a fast internet connection. A multi-core or a cloud supercomputer instance is recommended for the best experience.

Dependencies: Please see official midisimx repo for detailed information.

1) Install

# Install easily with pip
pip install -U galaxymidi

2) Download

# Download the dataset from Hugging Face

# Please note that the dataset size is 152GB
# so it make take a while to download
# depending on your internet speed
import galaxymidi

galaxymidi.download_dataset()

Alternatively, download directly with the Hugging Face CLI:

pip install -U "huggingface_hub[cli]"
huggingface-cli download projectlosangeles/Galaxy-MIDI-Dataset --repo-type dataset --local-dir Galaxy-MIDI-Dataset

3) Extract

# Extract fast as follows
from galaxymidi import fast_parallel_extract

fast_parallel_extract.fast_parallel_extract()

4) Quickstart

import galaxymidi

# Load one of the curated file lists
all_midis_files_list = galaxymidi.read_jsonl()

import galaxymidi

# Load the pre-computed Galaxy embeddings (midisimx format)
embeddings = galaxymidi.load_embeddings()

import galaxymidi

# Load the pre-computed comprehensive MIDI features
features = galaxymidi.load_features()

import galaxymidi

# To audition the MIDIs, render them as such
output_wav_file = galaxymidi.render_midi('my_midi.mid')

---

Documentation

galaxymidi API Reference

galaxymidi API Functions Index

---

Dataset structure

Galaxy MIDI Dataset
├── ATTRIBUTION.txt
├── Artwork
│   ├── Galaxy-MIDI-Dataset-Artwork (1).png
│   ├── Galaxy-MIDI-Dataset-Artwork (2).png
│   ├── Galaxy-MIDI-Dataset-Artwork (3).png
│   ├── Galaxy-MIDI-Dataset-Artwork (4).png
│   ├── Galaxy-MIDI-Dataset-Artwork (5).png
│   ├── Project-Los-Angeles.png
│   └── Tegridy-Code-2026.png
├── Code
│   ├── LICENSE
│   ├── midisimx-main.zip
│   └── tegridy-tools-main.zip
├── Embeddings
│   └── galaxy_midi_embeddings_1_2_1_2_weighted.bin
├── Features
│   └── galaxy_midi_features.ldmb
├── Files Lists
│   ├── aligned_midis_files_list.jsonl
│   ├── all_midis_files_list.jsonl
│   ├── clean_midis_files_list.jsonl
│   ├── drum_track_midis_files_list.jsonl
│   ├── genres_midis_files_list.jsonl
│   ├── karaoke_midis_files_list.jsonl
│   ├── mono_melodies_midis_files_list.jsonl
│   └── solo_melodies_midis_files_list.jsonl
├── GITHUB-REPO-LINK.txt
├── Genres
│   └── genres_data.jsonl
├── HUGGING-FACE-REPO-LINK.txt
├── Identified
│   └── identified_midis.jsonl
├── Karaoke
│   └── karaoke_midis_data.jsonl
├── LICENSE-CC-BY-NC-SA.txt
├── Lyrics
│   └── lyrics_midis.jsonl
├── MIDIs
│   └── ...
└── SoundFonts
│   ├── Expressive Flute SSO-v1.2.sf2
│   ├── KBH-Real-Choir-V2.5.sf2
│   ├── Nice-Strings-PlusOrchestra-v1.6.sf2
│   └── SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2
└── Technical Report
    ├── README.md
    ├── report.txt
    ├── stats_summary.json
    └── plots
        └── ...

12 root directories, 19893571 MIDI files, 32 data files

---

Dataset statistics

Please see Galaxy MIDI Dataset Technical Report for detailed stats and visualizations

• 19893571 unique, standardized, and original MIDI files
• 19893571 pre-computed comprehensive MIDI features for all MIDIs in the dataset
• 17292982 clean midis (melodic midis)
• 15591559 pre-computed MIDI embeddings
• 9361365 identified MIDI files
• 3667206 MIDIs with monophonic melodies
• 1355292 solo melody MIDI files
• 1227998 MIDIs aligned to bars
• 742915 solo drum track MIDI files
• 154990 Karaoke MIDI files
• 89569 MIDIs with genre, artist, and song labels
• 47069 MIDIs with corresponding English lyrics

---

Dataset integrity verification info

Main archive

• File: Galaxy-MIDI-Dataset.tar.gz
• SHA256: 0c43dc4ba8a2763aadb69a55d6fae617a718c1723283dacf0e8a44f11fe6cde8

---

MIDI Features

• File: galaxy_midi_features.ldmb
• SHA256: ae68d19facb311249204bec71fe6af96af18709de01591813697380147f72894

---

• File: galaxy_midi_dataset_19893571_midis_features_cc_by_nc_sa.ldmb
• SHA256: ae68d19facb311249204bec71fe6af96af18709de01591813697380147f72894

MIDI Embeddings

• File: galaxy_midi_embeddings_1_2_1_2_weighted.bin
• SHA256: 6f816f4f4bbf749e77b1e01ab1e56e76fe441db764140f4b438f9933af5cf726

---

• File: galaxy_midi_dataset_15591559_clean_midis_embeddings_1_2_1_2_weighted_cc_by_nc_sa.bin
• SHA256: 6f816f4f4bbf749e77b1e01ab1e56e76fe441db764140f4b438f9933af5cf726

---

• File: galaxy_midi_dataset_15591559_clean_midis_embeddings_1_2_1_2_weighted_cc_by_nc_sa.npy
• SHA256: be721de5ef866a784e976fa5a2b0de0c20adb23c3fd6ba5623fc70e5d42129ed

---

License

• The dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (see LICENSE-CC-BY-NC-SA.txt). In short: you may share and adapt the dataset for non-commercial purposes, provided you give appropriate credit (see ATTRIBUTION.txt and Citations) and distribute derivatives under the same license.
• Bundled code in the Code directory is distributed under its own license terms — see Code/LICENSE.
• The bundled SoundFonts are subject to their respective original licenses.

---

Citations

@misc{project_los_angeles_2026,
    author       = { Project Los Angeles and Tegridy Code },
    title        = { Galaxy MIDI Dataset },
    year         = 2026,
    url          = { https://huggingface.co/projectlosangeles/Galaxy-MIDI-Dataset },
    publisher    = { Hugging Face }
}

@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 }
}

@misc{TegridyMIDIDataset2025,
  title        = {Tegridy MIDI Dataset: Ultimate Multi-Instrumental MIDI Dataset for MIR and Music AI purposes},
  author       = {Alex Lev},
  publisher    = {Project Los Angeles / Tegridy Code},
  year         = {2025},
  url          = {https://github.com/asigalov61/Tegridy-MIDI-Dataset}
}

@misc {breadai_2025,
    author       = { {BreadAi} },
    title        = { Sourdough-midi-dataset (Revision cd19431) },
    year         = 2025,
    url          = {\url{https://huggingface.co/datasets/BreadAi/Sourdough-midi-dataset}},
    doi          = { 10.57967/hf/4743 },
    publisher    = { Hugging Face }
}

@inproceedings{bradshawaria,
  title={Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling},
  author={Bradshaw, Louis and Colton, Simon},
  booktitle={International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=X5hrhgndxW}, 
}

---

Project Los Angeles

Tegridy Code 2026