MuSP-Bench

490 human-authored questions across musical scores and performances

Paper · Website

MuSP-Bench assesses analytical and interpretive reasoning across 18 classical piano pieces or movements and six orchestral excerpts. It provides ABC notation, score images, MIDI-as-text, audio, and combined inputs, with questions spanning short passages and longer musical contexts.

The benchmark

Questions are annotated by modality, evidence horizon, content, and operation. S/P results separate tonality/style recognition and composer/title identification from all remaining questions (General).

Subset	Evidence	Questions
S	Score	180
P	Performance	106
S&P	Score and performance jointly	72
S/P — General	Either	36
S/P — Tonality/style	Either	48
S/P — Composer/title	Either	48

Evidence horizons are Short (1–2 bars or roughly under 5 seconds; 94 questions), Mid (3–16 bars or roughly 5–30 seconds; 123), Long (larger or noncontiguous spans; 177), and Any (96 contextual questions).

The 24 selections contain 227 score pages and approximately 110 minutes of audio. Recording lengths range from 17 seconds to 12 minutes 29 seconds.

Each plot counts selections, using one-page and one-minute bins. Per-piece inventory.

Load the data

from pathlib import Path
from datasets import load_dataset
from huggingface_hub import snapshot_download

repo = "bryel-labs/MuSP-Bench"
questions = load_dataset(repo, "default", split="test")
root = Path(snapshot_download(repo_id=repo, repo_type="dataset"))

There is one test split. Each row includes question_id, question, modality, horizon, category, operation, accepted answers, and input paths. Paths are relative to the downloaded repository; image contains an ordered list of page paths.

Twenty-one recordings are bundled. For the three external recordings, audio is null: download the BSED archive and use the file named in audio_archive_member. The dataset also provides audio_source_url.

Inputs and preprocessing

Input	Dataset field	Description
ABC	abc	Written score
Score images	image	Ordered score pages
PDF	pdf	Score document
Audio	audio	Selected recording
MIDI-as-text v1	midi1	Performance events with scientific pitch names
MIDI-as-text v2	midi2	Alternative serialization with numeric pitches
ABC+MIDI	abc_midi	ABC with aligned timing, velocity, and pedal information

For image+audio and ABC+audio, pair the corresponding inputs. S uses ABC or images; P uses MIDI-as-text or audio; S&P uses combined inputs; S/P uses any of the four individual formats.

Piano scores come from ASAP/(n)ASAP and were cleaned using consensus evidence from PianoCoRe. Selected piano recordings and performance MIDI originate from MAESTRO through ASAP. Orchestral scores, recordings, and note alignments come from BSED.

Composer/title information is withheld from music-input prompts. Supply only the question, required answer-format instructions, and selected input; accepted answers and reference examples are not prompt inputs. The metadata control supplies composer and title instead of musical evidence.

Orchestral MIDI-as-text uses recording-specific timing and score-derived velocities. ABC+MIDI retains its separately prepared alignment. MIDI v1 and v2 differ in more than pitch notation: headers, ordering, and controller details can also differ. These are benchmark serialization names, not MIDI-standard versions.

PianoCoRe score cleaning

PianoCoRe provides multiple MIDI performances of the same work, allowing recurring disagreements with the source score to be checked across recordings. This helps distinguish likely score errors from a single performer's missed or altered note.

The cleaning workflow aligns the performances to the score and gathers evidence for each candidate note correction. Notes repeatedly left unmatched are flagged for examination; nearby performed pitches are then compared across recordings to identify supported replacements. The search follows each performance's local timing, accommodating tempo differences rather than comparing fixed timestamps across recordings.

Consensus evidence guides note-level corrections while preserving the score's written context. Tied continuations are treated separately because they do not require a new performed attack, and pitch corrections retain appropriate enharmonic spelling.

Format examples

These excerpts use the same selection (piece 01). Text examples show only the opening notes or first bar of the first voice; links provide the full inputs.

ABC (full score):

(3DEF (3GFE (3FGA (3BAG A2d2^c2=c2 |

MIDI-as-text v1 (full input):

D4 v33 00:00.50-00:00.65
E4 v43 00:00.68-00:00.82
F4 v42 00:00.85-00:01.01

MIDI-as-text v2 (full input):

62 v33 00:00.50-00:00.65
64 v43 00:00.68-00:00.82
65 v42 00:00.85-00:01.01

Both list pitch, velocity (v), onset, and end time. For example, D4 in v1 is MIDI pitch 62 in v2.

ABC+MIDI (full input):

(3DEF (3GFE (3FGA (3BAG A2d2^c2=c2 |
%%on   00:00.50 00:00.68 00:00.85 00:01.03 00:01.20 00:01.37 00:01.55 00:01.71 00:01.89 00:02.06 00:02.23 00:02.39 00:02.58 00:03.19 00:03.79 00:04.37 |
%%off  00:00.65 00:00.82 00:01.01 00:01.21 00:01.35 00:01.51 00:01.69 00:01.91 00:02.09 00:02.28 00:02.42 00:02.58 00:02.91 00:03.85 00:04.44 00:04.97 |
%%vel  33 43 42 44 45 56 47 62 67 69 74 65 70 62 65 61 |
%%ped  ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? |

The aligned rows give onset (on), end (off), velocity (vel), and pedal (ped) information; ? denotes unavailable pedal information.

Score image / PDF — first page below; open the PDF.

Audio — listen to the recording.

Image+audio pairs the score pages above with this recording; ABC+audio pairs the linked ABC score with the same recording.

Results

Accuracy uses task-specific deterministic answer normalization, without partial credit or automatic timestamp tolerance.

Model	Weighted max accuracy
GPT-5.6 Sol	56.53%
Qwen3.5-Omni-Plus	23.06%
Muse Spark 1.2	44.08%

Weighted max takes the best evaluated format within each subset and weights by question count, excluding metadata and v2. It summarizes the best observed formats rather than one fixed configuration. AF-Next-Instruct has partial subset coverage, so its results appear in the full results without a full-benchmark aggregate.

MIDI serialization ablation

The comparison covers 238 P and S/P questions per model.

Model	v1 accuracy	v2 accuracy	Drop (percentage points)
GPT-5.6 Sol	56.30%	31.93%	24.4
Qwen3.5-Omni-Plus	17.65%	10.08%	7.6
Muse Spark 1.2	38.24%	37.82%	0.4

Drops use unrounded counts and reflect the complete serialization change. Detailed ablation results.

Two saved responses use earlier question wording and are marked in the predictions. Current-prompt results exclude them and report adjusted coverage.

To rescore saved predictions with Node.js 20 or newer:

node evaluation/evaluate.mjs --predictions results/predictions.jsonl

Use and attribution

MuSP-Bench is a test benchmark focused on Western classical music. Interpretive answers can admit alternatives beyond the references, and questions from the same piece are related. Public source music may overlap with model training data.

The release is CC BY-NC-SA 4.0, subject to source-specific terms. Credit MuSP-Bench and its source datasets. The three external Blomstedt recordings are not redistributed. See LICENSE.md.

Citation

@misc{liessensdujardin2026muspbench,
  title = {MuSP-Bench: Advanced Multimodal Benchmarking of Music Understanding Across Score and Performance},
  author = {Liessens Dujardin, Milan and Yu, Song-Ze and Miao, Kevin},
  year = {2026},
  url = {https://arxiv.org/abs/2608.28212},
  note = {Benchmark and dataset release}
}