projectlosangeles/Monster-MIDI-Dataset

Monster MIDI Dataset

Giant searchable raw MIDI dataset for MIR and Music AI purposes

---

Monster MIDI Dataset GPU Search and Filter

[NEW] Monster GPU/CPU Search and Filter stand-alone Python module with improved matching

Installation

Install requirements

!git clone --depth 1 https://github.com/asigalov61/Monster-MIDI-Dataset

!pip install cupy-cuda12x
!pip install numpy==1.24.4
!pip install huggingface_hub
!pip install hf-transfer
!pip install ipywidgets
!pip install tqdm

Import modules

import os
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"

from huggingface_hub import hf_hub_download

%cd ./Monster-MIDI-Dataset
import monster_search_and_filter
%cd ..

Download and unzip Monster MIDI dataset

hf_hub_download(repo_id='projectlosangeles/Monster-MIDI-Dataset',
                repo_type='dataset',
                filename='Monster-MIDI-Dataset-Ver-1-0-CC-BY-NC-SA.zip',
                local_dir='./Monster-MIDI-Dataset/'
                )

%cd ./Monster-MIDI-Dataset
!unzip -o Monster-MIDI-Dataset-Ver-1-0-CC-BY-NC-SA.zip > /dev/null
%cd ..

Run the search

sigs_data_path = './Monster-MIDI-Dataset/SIGNATURES_DATA/MONSTER_SIGNATURES_DATA.pickle'

sigs_data = monster_search_and_filter.load_pickle(sigs_data_path)
sigs_dicts = monster_search_and_filter.load_signatures(sigs_data)
X, global_union = monster_search_and_filter.precompute_signatures(sigs_dicts)

monster_search_and_filter.search_and_filter(sigs_dicts, X, global_union)

[LEGACY]

Search, filter and explore Monster MIDI Dataset :)

PLEASE NOTE: Google Colab Pro or Pro+ subscription/A100 GPU is required to use the provided colab/code because of the size of the dataset and its data files

---

Monster MIDI Dataset Sample Search Results

Here are the Monster MIDI Dataset Sample Search Results

It takes about one hour on A100 GPU to do a full search on 285 source MIDIs

Please also check out Quad Music Transformer which was trained using these sample search results

---

Monster Music Transformer

Here is the large model trained on the full Monster MIDI Dataset to demo the dataset in action :)

Model was trained on full Monster MIDI Dataset for 65 hours (1 epoch) @ 4 batches on a single H100 GPU

This model can be used for music generation/composition or for (dataset) embeddings exploration

---

Enjoy and please CC BY-NC-SA :)

---

Project Los Angeles

Tegridy Code 2025