README.md

---
pretty_name: MuSP-Bench
language:
- en
license: cc-by-nc-sa-4.0
task_categories:
- question-answering
configs:
- config_name: default
  data_files:
  - split: test
    path: data/questions.csv
- config_name: ablation
  data_files:
  - split: test
    path: data/ablation.csv
---

# MuSP-Bench

MuSP-Bench is a 490-question benchmark for musical score understanding,
performance listening, and combined score-performance reasoning.

## Contents

- `data/questions.csv`: all 490 questions, accepted answers, and the
  response contract for each.
- `data/ablation.csv`: MIDI2 baseline accuracy by model and reporting subset.
- `inputs/pdf/without_context/`: one context-removed PDF per piece.
- `inputs/images/`: rendered score-page images for every piece.
- `inputs/abc/`: one ABC score per piece.
- `inputs/abc_plus_midi/`: one aligned ABC+MIDI score-performance representation per piece.
- `inputs/midi_performance_event_text/`: one MIDI performance-event-text file per piece.
- `inputs/midi2_performance_event_text/`: one MIDI2 numeric-pitch
  performance-event-text file per piece.
- `inputs/audio/`: WAV performances for 21 pieces. The three Blomstedt
  performances remain in the original BSED release and are linked below.

Each question has associated audio, MIDI and MIDI2 performance event text,
ABC, ABC+MIDI, score images, and a metadata-removed PDF. The `audio` field is a
relative local path for bundled recordings and an official source URL for the
three externally hosted BSED recordings.

## Input baselines

- **Metadata** supplies contextual information without musical evidence.
- **ABC** represents the written score as symbolic text, while **Image** uses
  rendered score pages and **Audio** uses the performance recording.
- **MIDI-as-text (MIDI1)** is a chronological performance-event list containing
  named pitches, attack velocities, onset times, and acoustic end times.
  **MIDI2** replaces pitch names with MIDI note numbers.
- **ABC+MIDI** preserves the ABC score and aligns each written note with its
  performed onset, offset, velocity, and pedal state using parallel `%%on`,
  `%%off`, `%%vel`, and `%%ped` rows.
- **Image+audio** and **ABC+audio** pair the corresponding score
  representation with the recording.

## Ablation

The `ablation` configuration compares MIDI1, which uses named pitches, with
MIDI2, which replaces them with numeric MIDI note values. The numeric encoding
substantially reduces overall accuracy for GPT-5.6-sol (55.88% to 29.41%) and
both Qwen models, while MuseSpark remains at 34.45%.

| Representation | Model | P | General (S/P) | Tonality/style (S/P) | Composer/title (S/P) | Overall |
|---|---|---:|---:|---:|---:|---:|
| MIDI1 | GPT-5.6-sol | 63/106 (59.43%) | 18/36 (50.00%) | 37/48 (77.08%) | 15/48 (31.25%) | 133/238 (55.88%) |
| MIDI1 | MuseSpark 1.2 | 28/106 (26.42%) | 11/36 (30.56%) | 32/48 (66.67%) | 11/48 (22.92%) | 82/238 (34.45%) |
| MIDI1 | Qwen3.5-Omni-Plus | 12/106 (11.32%) | 5/36 (13.89%) | 19/48 (39.58%) | 0/48 (0.00%) | 36/238 (15.13%) |
| MIDI1 | Qwen3.6-Plus | 11/106 (10.38%) | 4/36 (11.11%) | 17/48 (35.42%) | 0/48 (0.00%) | 32/238 (13.45%) |
| MIDI2 | GPT-5.6-sol | 34/106 (32.08%) | 11/36 (30.56%) | 25/48 (52.08%) | 0/48 (0.00%) | 70/238 (29.41%) |
| MIDI2 | MuseSpark 1.2 | 28/106 (26.42%) | 11/36 (30.56%) | 32/48 (66.67%) | 11/48 (22.92%) | 82/238 (34.45%) |
| MIDI2 | Qwen3.5-Omni-Plus | 9/106 (8.49%) | 3/36 (8.33%) | 9/48 (18.75%) | 0/48 (0.00%) | 21/238 (8.82%) |
| MIDI2 | Qwen3.6-Plus | 4/106 (3.77%) | 1/36 (2.78%) | 9/48 (18.75%) | 0/48 (0.00%) | 14/238 (5.88%) |

## IDs and ordering

Rows are ordered by modality, composer, and piece number, then by their
underlying span-specific horizon and question order.

For S/P questions, `submodality` is `general`, `tonality/style`, or
`composer/title`, following the benchmark's established reporting split. It is
blank for S, P, and SP questions.

## Horizons

- `Short`: one event, up to two measures, or approximately 10 seconds.
- `Mid`: 3–16 measures or approximately 10–60 seconds.
- `Long`: more than 16 measures, a section, a whole-piece search, or a
  nonlocal comparison.
- `Any`: the question is not constrained to a particular musical span. This is
  used for S/P `tonality/style` and `composer/title` questions.

## Modalities

- `S`: answer from the score.
- `P`: answer from the performance recording.
- `SP`: answer using both score and performance.
- `S/P`: answer from either source.

`answerExample` contains JSON-encoded formatting examples. Examples illustrate
the requested syntax only and are not answers to the benchmark questions.
`answers` contains the JSON-encoded list of accepted ground-truth answers.

## Sources and attribution

MuSP-Bench is derived from the following source datasets. The released inputs
have been selected, cropped or normalized as needed, converted into additional
representations, and paired with newly authored benchmark questions and
accepted answers.

- **ASAP — Aligned Scores and Performances** (pieces 1–18): Francesco
  Foscarin, Andrew McLeod, Philippe Rigaux, Florent Jacquemard, and Masahiko
  Sakai, *ASAP: a dataset of aligned scores and performances for piano
  transcription*, ISMIR 2020.
  [Dataset](https://github.com/fosfrancesco/asap-dataset) ·
  [license](https://github.com/fosfrancesco/asap-dataset/blob/master/LICENSE.md)
- **MAESTRO v2.0.0**: Curtis Hawthorne, Andriy Stasyuk, Adam Roberts, Ian
  Simon, Cheng-Zhi Anna Huang, Sander Dieleman, Erich Elsen, Jesse Engel, and
  Douglas Eck, *Enabling Factorized Piano Music Modeling and Generation with
  the MAESTRO Dataset*, ICLR 2019. MAESTRO supplies the underlying competition
  recordings used by the selected ASAP performances.
  [Dataset and citation](https://magenta.withgoogle.com/datasets/maestro)
- **PianoCoRe**: Ilya Borovik, *PianoCoRe: Combined and Refined Piano MIDI
  Dataset*, TISMIR 2026. PianoCoRe refined alignments were used where available
  in preparing the piano score/performance representations.
  [Dataset, citation, and license](https://huggingface.co/datasets/SyMuPe/PianoCoRe)
- **BSED — Beethoven Symphony Excerpt Dataset** (pieces 19–24): Hans-Ulrich
  Berendes, Abhirup Saha, Ben Maman, Vlora Arifi-Müller, and Meinard Müller,
  *Beethoven Symphony Excerpt Dataset (BSED): An Evaluation Dataset for
  Orchestral Music Transcription*, TISMIR 2026.
  [Dataset and citation](https://zenodo.org/records/20344500)

BSED documents that its first three real-audio splits use public-domain IMSLP
recordings, while its fourth split uses short excerpts from Herbert Blomstedt
CD recordings that are not public domain; those excerpts are distributed by
the BSED authors under European scientific quotation and citation provisions.
See the [source paper's ethical statement](https://doi.org/10.5334/tismir.343)
for details.

The Blomstedt audio is not redistributed in this repository. Download the
[official BSED archive](https://zenodo.org/records/20344500/files/BSED.zip?download=1)
and use these archive-internal files:

- Piece 19: `02_Audio/wav_44100/BSED-04_4_Beethoven_Op036-01_Blomstedt1978.wav`
- Piece 20: `02_Audio/wav_44100/BSED-05_4_Beethoven_Op036-03_Blomstedt1978.wav`
- Piece 23: `02_Audio/wav_44100/BSED-18_4_Beethoven_Op093-02_Blomstedt1978.wav`

## License

MuSP-Bench is released under the
[Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
license](https://creativecommons.org/licenses/by-nc-sa/4.0/). Credit the
MuSP-Bench authors and the applicable source datasets above, link the license,
and indicate modifications. Underlying third-party rights and applicable
copyright exceptions remain unaffected.