Awesome

LLark: A Multimodal Instruction-Following Language Model for Music

This is the code release associated with the paper:

LLark: A Multimodal Instruction-Following Language Model for Music<br /> Josh Gardner, Simon Durand, Daniel Stoller, Rachel Bittner.<br /> Published at ICML 2024

This repository contains the code used to build the training dataset, preprocess existing open-source music datasets, train the model, and run inference. Note that this paper is not accompanied with any trained models.

For more details about this work, you can read the preprint of the paper, read the Spotify Research blog post about this work, and listen to demos on the companion site. See also the ICML page for the paper.

This is not an officially supported Spotify product.

Citing This Repository

If you use this code in a publication, please cite our paper:

@article{gardner2023llark,
  title={LLark: A Multimodal Instruction-Following Language Model for Music},
  author={Gardner, Josh and Durand, Simon and Stoller, Daniel and Bittner, Rachel},
  journal={Proc. of the International Conference on Machine Learning (ICML)},
  year={2024}
}

Using The Code

Docker Environments

All of the code in this repo should be run in one of the provided Docker containers (see docker). There are three Dockerfiles for separate use cases:

• m2t-train.dockerfile is for model training and inference
• m2t-preprocess.dockerfile is for data preprocessing (i.e. running Beam pipelines)
• jukebox-embed.dockerfile is for extracting Jukebox embeddings

Warning

Note that the data preprocessing pipeline can be run locally or on Google Cloud Dataflow, a cloud-based data processing system that allows the code to be sped up by running in parallel across potentially thousands of machines. Using these pipelines on Dataflow may incur costs against your own Google Cloud project.

Running these pipelines on Google Cloud Dataflow may also require pushing one or more Docker images to Google Artifact Registry, properly setting up permissions to use these Docker images from within Google Cloud Dataflow, and replacing various hard-coded Docker image paths with paths to your uploaded Docker images. This repository contains no code to aid with this process.

Preprocess Your Own Audio

We provide some utilitities for converting, cropping, and annotating audio datasets using the process described in our paper. These utilities are located in scripts/preprocessing; see the header of each file in that directory for usage examples.

All of the preprocessing scripts use Apache Beam (on Google DataFlow). In order to run these code examples, we recommend using the Docker environment defined in docker/m2t-preprocess.dockerfile.

Generate Instruction-Tuning Data

The file scripts/openai/fetch_openai_instruct_data.py can be adapted to generate instruction-tuning data. The prompts for each dataset are located in the m2t/instruct directory.

Extract Jukebox Embeddings

The Jukebox embedding pipeline uses Apache Beam (on Google Cloud Dataflow). In order to run these code examples, we recommend using the Docker environment defined in docker/jukebox-embed.dockerfile. Embedding a set of around 100k audio files takes less than 1 hour using the default parameters in that script.

For using CLAP, we provide a similar set of utilities in the scripts/clap subdirectory; the CLAP embedding script can be executed from the m2t-preprocess.dockerfile Docker environment and does not require its own separate Docker environment.

Training

This repo does not officially support training. However, we provide scripts that could be adapted to train a model (mostly as a way to describe the exact training hyperparameters used in training) in the scripts/training subdirectory. This includes the main LLark model, along with models based on MPT-1B and CLAP.

Evaluation

For reproducibility, we provide a set of notebooks that define the evaluation procedure used in our paper. These notebooks can be used or modified to support evaluation of other models. Note that in some cases, the notebooks require obtaining the original evaluation datasets (MusicNet, MedleyDB, GTZAN, Giant Steps (tempo, key)).

Contributing

We feel that a welcoming community is important and we ask that you follow Spotify's Open Source Code of Conduct in all interactions with the community.

Authors

• Josh Gardner jpgard@cs.washington.edu
• Peter Sobot psobot@spotify.com

Follow @SpotifyResearch on Twitter for updates.

License

Copyright 2023 Spotify, Inc.

Licensed under the Apache License, Version 2.0: https://www.apache.org/licenses/LICENSE-2.0

The m2t/llava directory contains a subset of the code from LLaVA. We do not make substantial modifications to this code but include it in order to make the repository self-contained and remove other external dependencies.

The jukebox directory contains code adapted from jukemir.

Security Issues?

Please report sensitive security issues via Spotify's bug-bounty program (https://hackerone.com/spotify) rather than GitHub.