---
license: cc0-1.0
task_categories:
- audio-classification
- text-generation
tags:
- music
- transcription
- arrangement
- musescore
- musicxml
- score
- piano
- orchestra
pretty_name: Operation Legato
size_categories:
- 10K<n<100K
dataset_info:
  features:
  - name: score_id
    dtype: string
  - name: song_name
    dtype: string
  - name: artist_name
    dtype: string
  - name: title
    dtype: string
  - name: arrangement_description
    dtype: string
  - name: complexity
    dtype: string
  - name: genres
    dtype: string
  - name: instruments
    dtype: string
  - name: instrumentations
    dtype: string
  - name: tags
    dtype: string
  - name: original_description
    dtype: string
  - name: n_tracks
    dtype: int64
  - name: midi_programs
    dtype: string
  - name: rating
    dtype: float64
  - name: n_views
    dtype: int64
  - name: n_favorites
    dtype: int64
  - name: score_duration_seconds
    dtype: float64
  - name: n_notes
    dtype: int64
  - name: has_lyrics
    dtype: bool
  - name: has_annotations
    dtype: bool
  - name: has_matched_audio
    dtype: bool
  - name: youtube_id
    dtype: string
  - name: youtube_title
    dtype: string
  - name: youtube_query
    dtype: string
  - name: audio_duration_seconds
    dtype: float64
  - name: alignment_confidence
    dtype: float64
  - name: time_stretch_ratio
    dtype: float64
  - name: tokens
    dtype: string
  splits:
  - name: train
    num_bytes: 3389708622
    num_examples: 71337
  - name: validation
    num_bytes: 201828567
    num_examples: 3963
  - name: test
    num_bytes: 188795634
    num_examples: 3963
  download_size: 596054938
  dataset_size: 3780332823
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# Operation Legato

A comprehensive dataset of 22,060 music score arrangements from MuseScore.

## Dataset Description

Each record contains:
- A **tokenized music score** in a simplified, absolute-timing text format
- **Arrangement metadata**: instrument(s), complexity level, genre, tags
- **YouTube reference** for the original song audio (when available)
- **Alignment information** between the score and original audio

### Source
Built from the [PDMX dataset](https://zenodo.org/records/13763756) - 254K public domain MusicXML scores scraped from MuseScore.com, filtered to the top 15,000 most popular songs with identifiable artists.

## Token Format

Scores are represented in a human-readable token format with absolute timing in seconds:

```
@meta complexity=intermediate
@meta genre=classical
@meta instruments=Piano
@meta arrangement=Solo
@time_sig t=0.000 4/4
@key_sig t=0.000 key=C mode=major
@tempo t=0.000 bpm=120.0
@annotation t=0.000 type=Dynamic mf
@note t=0.000 dur=0.500 p=60 v=80 ch=0 inst=Piano
@note t=0.500 dur=0.250 p=64 v=70 ch=0 inst=Piano
```

### Token types:
| Token | Description |
|-------|-------------|
| `@meta` | Score metadata (complexity, genre, instruments, arrangement type) |
| `@time_sig` | Time signature change with absolute timestamp |
| `@key_sig` | Key signature with key name and mode |
| `@tempo` | Tempo marking (BPM) with timestamp |
| `@tempo_text` | Textual tempo marking (e.g., "Allegro") |
| `@annotation` | Dynamics, articulations, section markings |
| `@note` | Note event: `t`=time(s), `dur`=duration(s), `p`=MIDI pitch, `v`=velocity, `ch`=channel, `inst`=instrument name |

## Fields

| Field | Type | Description |
|-------|------|-------------|
| `score_id` | string | Unique identifier (PDMX IPFS hash) |
| `song_name` | string | Original song name |
| `artist_name` | string | Original artist/composer |
| `title` | string | MuseScore arrangement title |
| `arrangement_description` | string | Human-readable description (e.g., "intermediate Piano Solo arrangement") |
| `complexity` | string | easy / intermediate / advanced / expert |
| `genres` | string | Genre(s) |
| `instruments` | string | Pipe-separated instrument list |
| `instrumentations` | string | Arrangement type (Solo, Duet, etc.) |
| `tags` | string | User-provided tags |
| `n_tracks` | int | Number of instrument tracks |
| `midi_programs` | string | MIDI program numbers per track |
| `rating` | float | Average user rating (0-5) |
| `n_views` | int | View count on MuseScore |
| `score_duration_seconds` | float | Score duration in seconds |
| `n_notes` | int | Total number of notes |
| `has_lyrics` | bool | Whether the score contains lyrics |
| `has_annotations` | bool | Whether the score contains dynamics/articulations |
| `has_matched_audio` | bool | Whether YouTube audio was found |
| `youtube_id` | string | YouTube video ID (if matched) |
| `youtube_query` | string | Search query to find the original audio |
| `alignment_confidence` | float | Score-to-audio alignment confidence (0-1) |
| `time_stretch_ratio` | float | Audio duration / score duration ratio |
| `tokens` | string | Full tokenized score content |

## Statistics

- **Total arrangements**: 22,060
- **With matched audio**: ~4,370 (growing)
- **Complexity**: easy (75%), intermediate (12%), advanced (11%), expert (2%)
- **Top instruments**: Piano, Flute, Bass guitar, Vocals, Drums, Cello, Saxophone, Violin, Trumpet, Clarinet
- **Genres**: classical (dominant), folk, soundtrack, rock, pop, jazz, electronic

## Use Cases

- Train music transcription models (audio -> score)
- Study arrangement patterns across instruments and difficulty levels
- Generate arrangements from audio in specific styles
- Music information retrieval with aligned score-audio pairs
- Symbolic music generation conditioned on instrument/difficulty

## Citation

Built from PDMX:
```bibtex
@inproceedings{long2024pdmx,
  title={PDMX: A Large-Scale Public Domain MusicXML Dataset for Symbolic Music Processing},
  author={Long, Phillip and Pati, Ashis},
  year={2024}
}
```