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

Dataset Structure

Each row in the dataset corresponds to a single JSON file from a user battle. The main fields are as follows:

• uuid (string): The unique ID for the battle.
• gateway_git_hash (string): The Git commit hash of the backend service that handled the request, useful for reproducibility.
• prompt (dict): Contains the user's original text prompt.
  • prompt (string): The text prompt provided by the user.
• prompt_detailed (dict): Structured information extracted from the prompt by the LLM router.
  • overall_prompt (string): The full text prompt.
  • instrumental (bool): Whether the prompt implies an instrumental track.
  • lyrics (string): User-provided lyrics, if any. (Currently unused)
  • duration (float): The duration in seconds requested by the user. Defaults to 0.0 if not specified. (Currently unused)
• prompt_user / vote_user (dict): Anonymized information about the user.
  • salted_ip (string): The user's salted and hashed IP address, allowing for longitudinal analysis without exposing personal data.
• prompt_session / vote_session (dict): Information about the user's session.
  • uuid (string): The unique ID for the user session.
  • create_time (float): The session creation timestamp.
  • frontend_git_hash (string): The Git commit hash of the frontend interface used during the session.
  • ack_tos (string): A hash representing the version of the Terms of Service acknowledged by the user.
  • new_battle_times (list of float): A list of timestamps indicating when new battles were initiated within the session.
• prompt_prebaked (bool): Indicates if the battle was pre-generated (true) rather than created in real-time by a user prompt.
• prompt_routed (bool): Indicates if the prompt was successfully processed by the LLM-based routing system.
• a_audio_url / b_audio_url (string): The original Google Cloud Storage URL for the audio file. This is retained for archival purposes.
• a_metadata / b_metadata (dict): Detailed metadata about the generated audio.
  • system_key (dict): Identifies the model used.
    • system_tag (string): The name of the model (e.g., acestep, sao).
    • variant_tag (string): The model variant used (e.g., initial, quick).
  • system_git_hash (string): The Git commit hash of the model endpoint service.
  • system_time_queued / started / completed (float): Timestamps for the generation process on the model server.
  • gateway_time_started / completed (float): Timestamps for the generation process as seen by the main gateway.
  • gateway_num_retries (int): The number of times the generation was retried.
  • size_bytes (int): The size of the generated audio file in bytes.
  • lyrics (string): The lyrics generated by the model (can be an empty string).
  • duration (float): The actual duration of the generated audio in seconds.
  • sample_rate (int): The sample rate of the audio (e.g., 44100).
  • num_channels (int): The number of audio channels (e.g., 1 for mono, 2 for stereo).
  • checksum (string): An MD5 checksum for data integrity verification.
• vote (dict): Contains all information related to the user's preference vote.
  • a_listen_data / b_listen_data (list of dict): A sequence of listening events for each track.
    • action (string): The user action (e.g., PLAY, PAUSE, TICK).
    • time (float): The timestamp of the action.
  • preference (string): The user's vote (A, B, TIE, or BOTH_BAD).
  • preference_time (float): The timestamp when the vote was cast.
  • feedback (string): General text feedback on the battle (can be an empty string).
  • a_feedback (string): Specific text feedback about audio A (can be an empty string).
  • b_feedback (string): Specific text feedback about audio B (can be an empty string).
  • feedback_time (float): The timestamp when the feedback was submitted.
• timings (list of dict): A sequence of events logging the backend generation process for performance analysis.
  • event (string): The name of the backend event (e.g., parse, generate_parallel_start).
  • time (float): The timestamp of the event.
• audio_a / audio_b (audio): The relative path to the audio file. It will be an empty string if the audio is not publicly released (e.g., for "sao" models).

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:

@misc{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},
    journal={arXiv:2507.20900},
    year={2025}
}

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.