zzsi/openscore — OpenScore Sheet Music Pages

Rendered sheet music pages from three open-score corpora, paired with per-page MusicXML ground truth. Intended for optical music recognition (OMR) research and supervised fine-tuning of vision-language models.

Images are rendered from source MusicXML via LilyPond (Emmentaler font). Per-page MusicXML is extracted by parsing bar numbers from the rendered SVGs and slicing the source score with music21.

---

Corpora

Corpus	Scores	Instrumentation	Source
lieder	~1,460	Voice + piano (3 staves)	OpenScore/Lieder
quartets	~122	String quartet (4 staves)	OpenScore/StringQuartets
orchestra	~94 movements	Full orchestra (10–20+ staves)	MarkGotham/Hauptstimme

---

Configs

pages_transcribed — image + per-page MusicXML (SFT-ready)

Each row is one rendered page paired with the MusicXML for the bars on that page. Suitable for supervised fine-tuning of OMR models.

Split	Rows
train	14,129
test	455
dev	329

Fields:

Field	Type	Description
image	PIL.Image	Full-page score render
score_id	str	Score identifier (e.g. lc6583477)
corpus	str	lieder, quartets, or orchestra
composer	str
opus	str
title	str
page	int	1-indexed page number
n_pages	int	Total pages in the score
bar_start	int	First bar number on this page
bar_end	int	Last bar number on this page (inclusive)
musicxml	str	MusicXML for bar_start–bar_end

---

pages — image only, all corpora

Same rows as pages_transcribed but without the musicxml, bar_start, and bar_end fields. Useful for unsupervised pre-training or image-only tasks.

Split	Rows
train	16,225
test	478
dev	339

---

scores — full MusicXML per score

One row per score (not per page). Contains the complete MusicXML for the entire piece plus metadata.

Split	Rows
train	1,424
test	79
dev	79

Fields: score_id, composer, opus, title, corpus, instruments (list), page (total pages), n_pages, musicxml (full score).

---

Usage

Load pages_transcribed

from datasets import load_dataset

ds = load_dataset("zzsi/openscore", "pages_transcribed")
example = ds["train"][0]
example["image"].show()
print(example["musicxml"][:500])

Filter by corpus (streaming)

The dataset is sorted by corpus within each split, so row groups in the parquet files are corpus-homogeneous. This means streaming with a corpus filter is efficient: non-matching row groups are skipped without being downloaded.

from datasets import load_dataset

# Lieder only
ds = load_dataset("zzsi/openscore", "pages_transcribed",
                  streaming=True, split="train")
ds = ds.filter(lambda r: r["corpus"] == "lieder")

# Lieder + quartets (no orchestra)
ds = ds.filter(lambda r: r["corpus"] in {"lieder", "quartets"})

Quick subset for testing

# First 100 rows (any corpus)
ds = load_dataset("zzsi/openscore", "pages_transcribed",
                  streaming=True, split="train")
sample = list(ds.take(100))

Fine-tuning example (Qwen-VL style)

from datasets import load_dataset

ds = load_dataset("zzsi/openscore", "pages_transcribed", split="train")

def to_chat(row):
    return {
        "messages": [
            {"role": "user", "content": [
                {"type": "image", "image": row["image"]},
                {"type": "text",  "text": "Transcribe this sheet music page to MusicXML."},
            ]},
            {"role": "assistant", "content": row["musicxml"]},
        ]
    }

ds = ds.map(to_chat)

---

Construction

1. Render: Source MusicXML is converted to LilyPond (.ly) format and rendered to SVG pages using a Docker image containing LilyPond 2.24. Bar numbers are made visible on every bar (all-bar-numbers-visible).
2. Align: Bar numbers are parsed from each SVG page to determine which bars appear on each page.
3. Slice: music21 slices the source MusicXML to the bar range for each page and re-exports it as a self-contained MusicXML fragment.

Pages whose bar numbers could not be reliably parsed (e.g. continuation pages with no bar number printed) are excluded.

---

Known Limitations

• Pickup bars: Scores with a pickup bar (anacrusis) have an implicit measure 0 that is accounted for in bar_start/bar_end.
• Orchestra page alignment: Orchestra scores frequently render to a different page count than the original due to \RemoveEmptyStaves in LilyPond. Alignment is based on bar numbers embedded in the rendered SVG, not on page index.
• MusicXML slice quality: Sliced MusicXML may be missing some cross-page spanners (slurs, hairpins). Inexpressible rhythms (rare) cause individual pages to be dropped.
• Render failures: ~6% of lieder scores, 3% of quartet scores, and 2 orchestra movements failed to render and are absent from the dataset.

---

License

Source scores are released under CC BY-SA 4.0. LilyPond renders and derived MusicXML slices carry the same license.

---

Attribution

• OpenScore Lieder — scores transcribed by the OpenScore community: https://github.com/OpenScore/Lieder
• OpenScore String Quartets — scores transcribed by the OpenScore community: https://github.com/OpenScore/StringQuartets
• Hauptstimme (Orchestra) — scores curated by Mark Gotham: https://github.com/MarkGotham/Hauptstimme

Rendering pipeline uses LilyPond and music21. Dataset construction code: https://github.com/zhudotexe/CVlization