music-arena/music-arena-dataset

Music Arena Dataset

This is the official dataset from Music Arena, an open platform for evaluating text-to-music (TTM) models.

How to Download (Recommended Method)

The most reliable way to get a complete local copy of all files, including the entire audio collection, is to clone the repository directly using Git. This method is ideal for offline access and workflows that require direct file manipulation.

Note: This repository uses Git LFS (Large File Storage) to manage the audio files. You must have Git LFS installed to download the audio files correctly.

# First, install Git LFS if you haven't already
# On Debian/Ubuntu: sudo apt-get install git-lfs
# On macOS: brew install git-lfs
git lfs install

# Clone the repository
git clone https://huggingface.co/datasets/music-arena/music-arena-dataset

⚠️ Note on Audio File Availability

Audio files generated by the following models are not included in this public data release due to specific release agreements:

• sao
• sao-small

The JSON metadata for battles involving these models is still included, but the audio_a and audio_b fields will be empty (null) for these specific files.

Dataset Structure

The dataset consists of a series of JSON files, each representing a single user battle. The data is stored in a flat (non-nested) structure for easy parsing.

Each row corresponds to one .json file with the following fields:

• battle_uuid (string): The unique ID for the battle.
• date (string): The ISO 8601 timestamp of when the battle occurred.
• prompt (string): The text prompt provided by the user.
• is_instrumental (bool): Whether the prompt was determined to imply an instrumental track.
• user_pseudonym (string): Anonymized user ID (salted and hashed IP) allowing for longitudinal analysis.
• audio_a (audio): The path to the audio file for model A. Will be null if the audio is not publicly released.
• system_a (string): The name of the model that generated audio_a.
• audio_b (audio): The path to the audio file for model B. Will be null if the audio is not publicly released.
• system_b (string): The name of the model that generated audio_b.
• preference (string): The user's vote, one of: "A", "B", "TIE", or "BOTH_BAD".
• total_listening_time_a (float): The total time in seconds the user spent listening to audio_a.
• total_listening_time_b (float): The total time in seconds the user spent listening to audio_b.
• feedback (string): Optional text feedback provided by the user (can be an empty string).

Platform and Resources

Music Arena is fully open-source. You can find more information through the links below:

• Music Arena: https://music-arena.org
• Source Code: https://github.com/gclef-cmu/music-arena
• Blog Post: https://gclef-cmu.org/blog/posts/250919_MusicArena/

Citation

If you use this dataset in your research, please cite the following paper:

@inproceedings{kim2025musicarena,
    title={Music Arena: Live Evaluation for Text-to-Music},
    author={Yonghyun Kim and Wayne Chi and Anastasios Angelopoulos and Wei-Lin Chiang and Koichi Saito and Shinji Watanabe and Yuki Mitsufuji and Chris Donahue},
    booktitle={Advances in Neural Information Processing Systems 38 (NeurIPS 2025)},
    year={2025},
    note={Creative AI Track},
    eprint={2507.20900},
    archiveprefix={arXiv}
}

Acknowledgements and Disclosure of Funding Music Arena is supported by funding from Sony AI, with informal and pro-bono assistance provided by LMArena. We extend our sincere thanks to our commercial contacts at Producer.ai (ex-Riffusion), Stability AI, Google DeepMind, and Suno for productive discussions that informed the key features and policies of Music Arena. Music Arena is approved by CMU's Institutional Review Board under Protocol STUDY2024_00000489.