MARI Dataset

Scripts for building the Music Add Remove Instruction (MARI) dataset. The dataset is used to instruction-tune and evaluate text-to-music models for ADD and REMOVE editing operations.

The dataset is built using the MoisesDB multitrack dataset. The dataset parquet file and the evaluation mixes are published on Hugging Face. The training mixes are not premixed and must be computed locally with compute_files.py.

Prerequisites

• uv (manages the Python 3.11+ environment automatically).
• The MARI dataset from Hugging Face, with parquet file and evaluation mixes.
• The MoisesDB multitrack dataset required to compute the training mixes. Not needed if you only evaluate using the provided evaluation mixes.

Setup

uv sync

Paths and parameters are found in config.py. By default:

• MOISES_FOLDER - ~/moises/songs: place your MoisesDB songs here.
• OUTPUT_FOLDER - ~/mari-dataset: where the parquet and mixes are read/written (~/mari-dataset/mari-dataset.parquet and ~/mari-dataset/songs/).

Either match this layout or edit config.py to point elsewhere.

Scripts

Script	Purpose
create_dataset.py	Create the parquet and mix all audio files from MoisesDB.
create_testset.py	Create the evaluation set and add the evaluation column.
compute_files.py	Compute the audio mixes referenced by the parquet. Use this when you have the parquet but not the mixes.
generate_instructions.py, audio_utils.py	Helpers for instruction text and audio mixing.

Run any script with uv run, e.g.:

uv run create_dataset.py             # create parquet + mixes from MoisesDB
uv run create_testset.py             # select the evaluation set
uv run compute_files.py              # compute all mixes from the parquet
uv run compute_files.py --eval-only  # compute only the evaluation mixes
uv run compute_files.py --force      # recompute even if files exist

Usage

Download the parquet file and the evaluation mixes from Hugging Face, then run compute_files.py to create the training mixes locally. Always use the published parquet and evaluation mixes so everyone evaluates against the exact same dataset.

create_dataset.py and create_testset.py are only used to create the dataset from MoisesDB and are not part of the normal workflow.

Authors

• Hugo Bachér: Main Author - KTH Royal Institute of Technology
• Mauro Luzzatto: Industry Collaborator - Epidemic Sound

Citing

If you use the MARI-dataset for your research, please cite the following:

@misc{bacher2026mari,
    title = {{MARI: Music Add Remove Instruction Dataset}},
    author = {Hugo Bachér and Mauro Luzzatto},
    year = {2026},
    month = jun,
    version = {1.0},
    url = {https://huggingface.co/datasets/Cactooz/MARI-dataset},
    note = {{Code available: \url{https://github.com/Cactooz/MARI-dataset}}}
}

License

MARI-dataset and MoisesDB are distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

For the complete license, see: https://creativecommons.org/licenses/by-nc-sa/4.0/