MSU-Bench: Musical Score Understanding Benchmark

Evaluating Large Language Models' Comprehension of Complete Musical Scores

---

Overview

MSU-Bench is a human-curated benchmark for evaluating the musical score understanding capabilities of Large Language Models (LLMs) and Vision-Language Models (VLMs). It supports multimodal evaluation through both textual (ABC notation) and visual (PDF/image) inputs.

Key Statistics:

• 150 complete musical scores
• 1,800 generative question-answer pairs
• 4 hierarchical difficulty levels
• 12 questions per score (3 per level)

Multi-level Understanding

Data Curation and Evaluation

Difficulty Levels

Level	Focus	Examples
Level 1 - Onset Information	Metadata at the beginning of a score	Composer, title, key, time signature, tempo, instrumentation, anacrusis
Level 2 - Notation & Note	Local bar-level notation details	Pitch range, accidentals, dynamics, articulations, ornaments, tempo changes
Level 3 - Chord & Harmony	Harmonic structures and progressions	Chord qualities, inversions, cadences, modulations, pedal points
Level 4 - Texture & Form	Large-scale structural analysis	Melodic motifs, thematic organisation, texture types, formal design

Dataset Configs

The dataset is available in two configurations:

flat (default) — 1 question per row, 1,800 rows

Each row contains a single question-answer pair with the score's ABC notation. Best for evaluation pipelines.

Column	Type	Description
song_id	string	Unique identifier derived from the score filename
abc_notation	string	Full ABC notation (text-based symbolic representation)
level	int32	Difficulty level (1–4)
question	string	The question text
answer	string	The reference answer

nested — 12 questions per score, 150 rows

Each row contains one complete score with all 12 questions nested, plus the PDF and page images. Best for per-score or visual analysis.

Column	Type	Description
song_id	string	Unique identifier
abc_notation	string	Full ABC notation
pdf	binary	The original rendered PDF
images	list[image]	Individual page images (PNG), 1–35 pages per score
questions	struct{level, question, answer}	12 questions (3 per difficulty level) with reference answers

Note: PDF and page images are only stored in the nested config to avoid duplication. Use song_id to join with flat if needed.

Modalities for Evaluation

Modality	Input	Target Models
Textual QA	ABC notation + question	LLMs
Visual QA	PDF/images + question	VLMs

Repertoire

The benchmark covers 150 scores from the Western art music canon, spanning:

• Periods: Baroque, Classical, Romantic, Impressionism, 20th Century
• Composers: Bach, Beethoven, Chopin, Brahms, Debussy, Liszt, Schubert, Mozart, Mussorgsky, Grieg, and others
• Genres: Sonatas, character pieces, fugues, waltzes, nocturnes, etudes, rhapsodies, symphonies, concertos, art songs

Usage

Loading the Flat Config (default)

from datasets import load_dataset

ds = load_dataset("Krinos/MSU-Bench", split="test")
print(len(ds))  # 1800

sample = ds[0]
print(sample["song_id"])
print(f"Level {sample['level']}: {sample['question']} -> {sample['answer']}")

Loading the Nested Config

ds_nested = load_dataset("Krinos/MSU-Bench", "nested", split="test")
print(len(ds_nested))  # 150

sample = ds_nested[0]
for lvl, q, a in zip(
    sample["questions"]["level"],
    sample["questions"]["question"],
    sample["questions"]["answer"],
):
    print(f"  L{lvl}: {q} -> {a}")

Visual QA with Page Images

ds_nested = load_dataset("Krinos/MSU-Bench", "nested", split="test")
sample = ds_nested[0]
for i, img in enumerate(sample["images"]):
    img.save(f"page_{i}.png")

Joining Flat + Nested for Visual Evaluation

ds_flat = load_dataset("Krinos/MSU-Bench", split="test")
ds_nested = load_dataset("Krinos/MSU-Bench", "nested", split="test")

# Build a lookup from song_id to images
images_lookup = {row["song_id"]: row["images"] for row in ds_nested}

# Get images for a flat row
sample = ds_flat[0]
images = images_lookup[sample["song_id"]]

Citation

@article{dai2025msubench,
  title={Musical Score Understanding Benchmark: Evaluating Large Language Models' Comprehension of Complete Musical Scores},
  author={Dai, Congren and Yang, Yue and Li, Krinos and Zhou, Huichi and Liang, Shijie and Zhang, Bo and Liu, Enyang and Jin, Ge and An, Hongran and Zhang, Haosen and Jing, Peiyuan and Lee, Kinhei and Zhang, Zhenxuan and Li, Xiaobing and Sun, Maosong},
  journal={arXiv preprint arXiv:2511.20697},
  year={2025}
}

License

Apache 2.0. Please also refer to the repository for licensing information regarding the musical scores sourced from MuseScore.

Acknowledgements

This work is supported by the Advanced Discipline Construction Project of Beijing Universities, the Special Programme of National Natural Science Foundation of China (Grant No. T2341003), and the Major Programme of National Social Science Fund of China (Grant No. 21ZD19).