README.md

---
license: cc-by-sa-4.0
task_categories:
- image-to-text
language:
- en
tags:
- music
- optical-music-recognition
- omr
- sheet-music
- musicxml
- lilypond
size_categories:
- 10K<n<100K
configs:
- config_name: pages
  data_files:
  - split: dev
    path: pages/dev-*
  - split: test
    path: pages/test-*
  - split: train
    path: pages/train-*
- config_name: pages-lieder
  data_files:
  - split: train
    path: pages-lieder/train-*
  - split: dev
    path: pages-lieder/dev-*
  - split: test
    path: pages-lieder/test-*
- config_name: pages_transcribed
  data_files:
  - split: dev
    path: pages_transcribed/dev-*
  - split: test
    path: pages_transcribed/test-*
  - split: train
    path: pages_transcribed/train-*
- config_name: scores
  data_files:
  - split: train
    path: scores/train-*
  - split: test
    path: scores/test-*
  - split: dev
    path: scores/dev-*
dataset_info:
- config_name: pages
  features:
  - name: image
    dtype: image
  - name: score_id
    dtype: string
  - name: corpus
    dtype: string
  - name: page
    dtype: int64
  - name: n_pages
    dtype: int64
  - name: composer
    dtype: string
  - name: opus
    dtype: string
  - name: title
    dtype: string
  splits:
  - name: dev
    num_bytes: 89590116
    num_examples: 331
  - name: test
    num_bytes: 100755142
    num_examples: 475
  - name: train
    num_bytes: 3981286117
    num_examples: 16165
  download_size: 3176298223
  dataset_size: 4171631375
- config_name: pages-lieder
  features:
  - name: score_id
    dtype: string
  - name: corpus
    dtype: string
  - name: page
    dtype: int64
  - name: n_pages
    dtype: int64
  - name: bar_start
    dtype: int64
  - name: bar_end
    dtype: int64
  - name: musicxml
    dtype: string
  splits:
  - name: train
    num_bytes: 511960929
    num_examples: 3415
  - name: dev
    num_bytes: 28971937
    num_examples: 195
  - name: test
    num_bytes: 31504305
    num_examples: 218
  download_size: 53578761
  dataset_size: 572437171
- config_name: pages_transcribed
  features:
  - name: score_id
    dtype: string
  - name: corpus
    dtype: string
  - name: page
    dtype: int64
  - name: n_pages
    dtype: int64
  - name: bar_start
    dtype: int64
  - name: bar_end
    dtype: int64
  - name: musicxml
    dtype: string
  - name: image
    dtype: image
  - name: composer
    dtype: string
  - name: opus
    dtype: string
  - name: title
    dtype: string
  splits:
  - name: dev
    num_bytes: 140682244
    num_examples: 327
  - name: test
    num_bytes: 156874375
    num_examples: 455
  - name: train
    num_bytes: 5585319264
    num_examples: 14111
  download_size: 3207002375
  dataset_size: 5882875883
- config_name: scores
  features:
  - name: score_id
    dtype: string
  - name: composer
    dtype: string
  - name: opus
    dtype: string
  - name: title
    dtype: string
  - name: corpus
    dtype: string
  - name: instruments
    list: string
  - name: page
    dtype: int64
  - name: n_pages
    dtype: int64
  - name: musicxml
    dtype: string
  splits:
  - name: train
    num_bytes: 1468576392
    num_examples: 1424
  - name: test
    num_bytes: 54739890
    num_examples: 79
  - name: dev
    num_bytes: 54992179
    num_examples: 79
  download_size: 154745773
  dataset_size: 1578308461
---

# 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](https://lilypond.org) (Emmentaler font). Per-page MusicXML is
extracted by parsing bar numbers from the rendered SVGs and slicing the source
score with [music21](https://web.mit.edu/music21/).

---

## Corpora

| Corpus | Scores | Instrumentation | Source |
|--------|-------:|-----------------|--------|
| `lieder` | ~1,460 | Voice + piano (3 staves) | [OpenScore/Lieder](https://github.com/OpenScore/Lieder) |
| `quartets` | ~122 | String quartet (4 staves) | [OpenScore/StringQuartets](https://github.com/OpenScore/StringQuartets) |
| `orchestra` | ~94 movements | Full orchestra (10–20+ staves) | [MarkGotham/Hauptstimme](https://github.com/MarkGotham/Hauptstimme) |

---

## ⚠️ Known issue — broken `musicxml` labels in the `quartets` corpus

In the `pages_transcribed` config, the per-page `musicxml` slicing
**fails on a large fraction of `quartets` pages**: the page image shows
real music, but the `musicxml` label is an empty stub (one measure per
part, `<attributes>` + end barline, **no notes**). Verified 2026-05-18
by a full scan of all splits.

| Corpus | Scores w/ empty pages | Empty pages | Genuine failed slices |
|--------|----------------------:|-------------|----------------------:|
| `quartets`  | 97 / 109   | 3,746 / 6,410 (58.4%) | 3,578 |
| `lieder`    | 13 / 1,018 | 70 / 3,392 (2.1%)     | 13 |
| `orchestra` | 14 / 91    | 251 / 5,111 (4.9%)    | 0 (spurious `bar_start > bar_end` pages) |

**Do not train on the `quartets` portion of `pages_transcribed`** until
the slicer is fixed — empty labels teach a model "page of music →
empty output". `lieder` and `orchestra` are usable (`lieder` has 13
stray failed-slice pages, listed below; `orchestra` has zero). The
`pages` config (images only) and `scores` config are unaffected. Key
signatures in non-empty labels are correct.

<details>
<summary>Broken quartet scores — <code>score_id (empty_pages/total_pages)</code> (97)</summary>

sq10307350 (14/26), sq10313029 (53/88), sq10328092 (8/16),
sq10372717 (12/25), sq10381459 (9/20), sq10406164 (53/128),
sq10414906 (13/22), sq10490761 (5/10), sq10517302 (116/198),
sq10527526 (67/115), sq10675759 (74/101), sq11154985 (79/119),
sq11164006 (134/204), sq11539384 (23/61), sq12113164 (21/34),
sq12536479 (50/115), sq12701461 (9/25), sq12772795 (9/25),
sq14720995 (12/25), sq15049456 (18/26), sq15230467 (14/19),
sq15624112 (38/60), sq15730717 (16/22), sq16138966 (49/79),
sq7103818 (46/62), sq7108150 (65/129), sq7114183 (9/21),
sq7123582 (52/88), sq7127785 (44/54), sq7224846 (54/68),
sq7249986 (21/34), sq7284122 (12/20), sq7294793 (33/48),
sq7295726 (63/81), sq7300376 (9/18), sq7302602 (47/59),
sq7302710 (33/65), sq7330550 (52/77), sq7353137 (16/28),
sq7354505 (166/299), sq7358579 (95/125), sq7358708 (22/40),
sq7383977 (14/38), sq7384409 (80/139), sq7397765 (63/161),
sq7434431 (11/17), sq7471661 (14/36), sq7483523 (118/156),
sq7524617 (15/33), sq7541288 (39/68), sq7551068 (17/26),
sq7556360 (81/118), sq7577795 (44/60), sq7588853 (25/38),
sq7872392 (79/127), sq8071278 (23/41), sq8075304 (9/18),
sq8088531 (10/22), sq8437280 (5/14), sq8437358 (4/11),
sq8438840 (2/9), sq8438915 (1/5), sq8438999 (2/10),
sq8454356 (42/68), sq8455808 (15/26), sq8461409 (9/13),
sq8509238 (17/26), sq8556926 (6/12), sq8561633 (42/62),
sq8623643 (64/104), sq8630159 (83/92), sq8796660 (13/20),
sq8806134 (10/14), sq8806746 (7/15), sq8806881 (9/13),
sq8807040 (13/23), sq8807667 (11/17), sq8811375 (74/102),
sq8818128 (68/109), sq8823783 (187/314), sq8853405 (12/31),
sq8885439 (79/175), sq8885571 (72/129), sq8907120 (18/33),
sq8913219 (46/61), sq8938822 (107/151), sq8940236 (21/57),
sq9010547 (50/62), sq9094235 (78/162), sq9137469 (26/52),
sq9146376 (9/17), sq9199617 (6/14), sq9396439 (16/24),
sq9529900 (70/80), sq9608209 (19/43), sq9719026 (41/55),
sq9961690 (15/28).

</details>

The 12 unaffected quartet scores: sq10502527, sq14387632, sq7070319,
sq7070781, sq7075297, sq7078259, sq7082029, sq7093885, sq7095930,
sq7236909, sq7648382, sq8812200.

Lieder scores with one stray failed-slice page each: lc6197282,
lc6486038, lc6593095, lc6613436, lc6613481, lc6614760, lc6624112,
lc6625925, lc6667483, lc6669339, lc6670960, lc6764425, lc8873154.

---

## 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`

```python
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.

```python
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

```python
# 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)

```python
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](https://creativecommons.org/licenses/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](https://lilypond.org) and
[music21](https://web.mit.edu/music21/). Dataset construction code:
https://github.com/zhudotexe/CVlization