Cactooz/MARI-dataset

MARI Dataset

MARI dataset for music instruction-tuning.

• Repository: https://github.com/Cactooz/MARI-dataset
• Language: English
• License: CC BY-NC-SA 4.0

Uses

Music Add Remove Instruction (MARI) dataset is a dataset for instruction-following music edits. The dataset is used to train and evaluate text-to-music models for ADD and REMOVE editing operations.

Dataset Structure

The mari-dataset.parquet has the following structure. Each row represents a pair of audio mixes, a small mix and a large mix, with the large mix containing additional instruments.

• song_id: UUID of the source song from MoisesDB.
• genre: Song genre from the source song from MoisesDB.
• small_stem: Stem category label of the smaller mix.
• large_stem: Stem category of the instruments that were added to go from the small to the large mix.
• small_file: Filename of the smaller mix.
• large_file: Filename of the larger mix.
• small_instrument_data: JSON list of {type, id} dicts for each instrument in the small mix.
• large_instrument_data: JSON list of {type, id} dicts for each instrument in the large mix.
• small_instruments: Human-readable comma-separated instrument names in the small mix.
• large_instruments: Human-readable comma-separated instrument names in the large mix.
• changed_instruments: Human-readable names of the instruments that were added (the delta between the small and large mix).
• add_instruction: Natural language instruction to ADD the changed instruments.
• remove_instruction: Natural language instruction to REMOVE the changed instruments.
• evaluation: Whether the sample is used for evaluation or not.

This dataset only have the input and target mixes for the evaluation set (evaluation: true) to ensure everyone uses the same dataset for evaluations. The evaluation set has 1000 ADD and 1000 REMOVE samples.

Training mixes need to be computed from MoisesDB separately using the scripts available on GitHub. MoisesDB can be downloaded from their website.

Dataset Creation

Source Data

The dataset was created using tracks from the MoisesDB multitrack dataset made by Music AI.

Annotations

Instrument annotations were made using MoisesDB original annotations and expanded with synonyms for the same instruments. Instruments that were part of ambiguous instrument types were labeled first using NVIDIA's Music Flamingo and then checked and tweaked manually.

Bias, Risks, and Limitations

Due to the underlying genre distribution in MoisesDB, the dataset is heavily biased toward rock, singer-songwriter, and pop songs.

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/