README.md

metadata

dataset_info:
  features:
    - name: identifier
      dtype: string
    - name: audio
      dtype: audio
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: distractor_1
      dtype: string
    - name: distractor_2
      dtype: string
    - name: distractor_3
      dtype: string
    - name: main_category
      dtype: string
    - name: secondary_categories
      dtype: string
    - name: difficulty
      dtype: string
    - name: excerpt_start_time
      dtype: string
    - name: excerpt_end_time
      dtype: string
    - name: song_link
      dtype: string
    - name: jamendo_id
      dtype: string
    - name: song_name
      dtype: string
    - name: artist_name
      dtype: string
    - name: album_name
      dtype: string
    - name: song_license_url
      dtype: string
  splits:
    - name: test
      num_bytes: 476422419
      num_examples: 320
  download_size: 476401140
  dataset_size: 476422419
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
license: cc-by-4.0
task_categories:
  - question-answering
language:
  - en
tags:
  - music
size_categories:
  - n<1K

HumMusQA: A Human-written Music Understanding QA Benchmark Dataset

Authors: Benno Weck, Pablo Puentes, Andrea Poltronieri, Satyajeet Prabhu, Dmitry Bogdanov

HumMusQA is a multiple-choice question answering dataset designed to test music understanding in Large Audio-Language Models (LALMs).

Dataset Highlights

• ✍️ 320 hand-written multiple-choice questions curated and validated by experts with musical training
• 🎵 108 Creative Commons-licensed music tracks sourced from Jamendo
• ⏱️ Music recordings ranging from 30 to 90 seconds

Loading the data

from datasets import load_dataset

dataset = load_dataset("mtg-upf/HumMusQA", split="test")

Licensing

• The dataset annotations are licensed under Creative Commons Attribution 4.0 (CC BY 4.0).
• Each audio track follows its own Creative Commons license, as specified in the dataset metadata.

Users are responsible for complying with the license terms of each individual audio track.

Paper

• Paper DOI: https://doi.org/10.18653/v1/2026.nlp4musa-1.9
• Zenodo DOI: https://doi.org/10.5281/zenodo.18462523
• arXiv DOI: https://doi.org/10.48550/arXiv.2603.27877

Citation

If you use this dataset, please cite our paper:

Benno Weck, Pablo Puentes, Andrea Poltronieri, Satyajeet Prabhu, and Dmitry Bogdanov. 2026. HumMusQA: A Human-written Music Understanding QA Benchmark Dataset. In Proceedings of the 4th Workshop on NLP for Music and Audio (NLP4MusA 2026), pages 58–67, Rabat, Morocco. Association for Computational Linguistics.

BibTeX

@inproceedings{weck-etal-2026-hummusqa,
    title = "{H}um{M}us{QA}: A Human-written Music Understanding {QA} Benchmark Dataset",
    author = "Weck, Benno  and
      Puentes, Pablo  and
      Poltronieri, Andrea  and
      Prabhu, Satyajeet  and
      Bogdanov, Dmitry",
    editor = "Epure, Elena V.  and
      Oramas, Sergio  and
      Doh, SeungHeon  and
      Ramoneda, Pedro  and
      Kruspe, Anna  and
      Sordo, Mohamed",
    booktitle = "Proceedings of the 4th Workshop on {NLP} for Music and Audio ({NLP}4{M}us{A} 2026)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.nlp4musa-1.9/",
    doi = "10.18653/v1/2026.nlp4musa-1.9",
    pages = "58--67",
    ISBN = "979-8-89176-369-2",
    abstract = "The evaluation of music understanding in Large Audio-Language Models (LALMs) requires a rigorously defined benchmark that truly tests whether models can perceive and interpret music, a standard that current data methodologies frequently fail to meet.This paper introduces a meticulously structured approach to music evaluation, proposing a new dataset of 320 hand-written questions curated and validated by experts with musical training, arguing that such focused, manual curation is superior for probing complex audio comprehension.To demonstrate the use of the dataset, we benchmark six state-of-the-art LALMs and additionally test their robustness to uni-modal shortcuts."
}