metadata

license: mit
tags:
  - symbolic-music
  - music-information-retrieval
  - classification
  - retrieval
  - benchmark

SyMuRBench Datasets and Precomputed Features

This repository contains datasets and precomputed features for SyMuRBench, a benchmark for symbolic music understanding models. It includes metadata and MIDI files for multiple classification and retrieval tasks, along with pre-extracted music21 and jSymbolic features.

You can install and use the full pipeline via: 👉 https://github.com/Mintas/SyMuRBench

Overview

SyMuRBench supports evaluation across diverse symbolic music tasks, including composer, genre, emotion, and instrument classification, as well as score-performance retrieval. This Hugging Face dataset provides:

Dataset metadata (CSV files)
MIDI files organized by task
Precomputed music21 and jSymbolic features
Configuration-ready structure for immediate use in benchmarking

Tasks Description

Task Name	Source Dataset	Task Type	# of Classes	# of Files	Default Metrics
ComposerClassificationASAP	ASAP	Multiclass Classification	7	197	weighted f1 score, balanced accuracy
GenreClassificationMMD	MetaMIDI	Multiclass Classification	7	2,795	weighted f1 score, balanced accuracy
GenreClassificationWMTX	WikiMT-X	Multiclass Classification	8	985	weighted f1 score, balanced accuracy
EmotionClassificationEMOPIA	Emopia	Multiclass Classification	4	191	weighted f1 score, balanced accuracy
EmotionClassificationMIREX	MIREX	Multiclass Classification	5	163	weighted f1 score, balanced accuracy
InstrumentDetectionMMD	MetaMIDI	Multilabel Classification	128	4,675	weighted f1 score
ScorePerformanceRetrievalASAP	ASAP	Retrieval	-	438 (219 pairs)	R@1, R@5, R@10, Median Rank

Precomputed Features

Precomputed features are available in the data/features/ folder:

music21_full_dataset.parquet
jsymbolic_full_dataset.parquet

Each file contains a unified table with:

midi_file: Filename of the MIDI
task: Corresponding task name
E_0 to E_N: Feature vector

Example

midi_file	task	E_0	E_1	...	E_672	E_673
Q1_0vLPYiPN7qY_1.mid	EmotionClassificationEMOPIA	0.0	0.0	...	0.0	0.0
Q1_4dXC1cC7crw_0.mid	EmotionClassificationEMOPIA	0.0	0.0	...	0.0	0.0

File Structure

The dataset is distributed as a ZIP archive:

data/datasets.zip

After extraction, the structure is:

datasets/
├── composer_and_retrieval_datasets/
│   ├── metadata_composer_dataset.csv
│   ├── metadata_retrieval_dataset.csv
│   └── ... (MIDI files organized in subfolders)
├── genre_dataset/
│   ├── metadata_genre_dataset.csv
│   └── midis/
├── wikimtx_dataset/
│   ├── metadata_wikimtx_dataset.csv
│   └── midis/
├── emopia_dataset/
│   ├── metadata_emopia_dataset.csv
│   └── midis/
├── mirex_dataset/
│   ├── metadata_mirex_dataset.csv
│   └── midis/
└── instrument_dataset/
    ├── metadata_instrument_dataset.csv
    └── midis/

CSV files: Contain filename and label (or pair info for retrieval).
MIDI files: Used as input for feature extractors.

How to Use

You can download and extract everything using the built-in utility:

from symurbench.utils import load_datasets

load_datasets(output_folder="./data", load_features=True)

This will:

Download datasets.zip and extract it
Optionally download precomputed features
Update config paths automatically

License

This dataset is released under the MIT License.

Citation

If you use SyMuRBench in your work, please cite:

@inproceedings{symurbench2025,
  author    = {Petr Strepetov and Dmitrii Kovalev},
  title     = {SyMuRBench: Benchmark for Symbolic Music Representations},
  booktitle = {Proceedings of the 3rd International Workshop on Multimedia Content Generation and Evaluation: New Methods and Practice (McGE '25)},
  year      = {2025},
  pages     = {9},
  publisher = {ACM},
  address   = {Dublin, Ireland},
  doi       = {10.1145/3746278.3759392}
}