README.md

metadata

license: cc-by-4.0
pretty_name: Strudel-Synth
task_categories:
  - text-generation
tags:
  - music
  - midi
  - strudel
  - symbolic-music
  - music-decompilation
  - code-generation
size_categories:
  - 10K<n<100K

Strudel-Synth

Strudel-Synth is a synthetic corpus of 21,174 (MIDI, Strudel) pairs for training and evaluating MIDI-to-Strudel decompilation, introduced in Decomposer: Learning to Decompile Symbolic Music to Programs.

• 🎹 Live demo: anonymousgiraffe/decomposer-demo
• 🤗 Model: anonymousgiraffe/Decomposer-Qwen3-8B

Each pair consists of a Strudel program distilled from Claude-Opus-4.6 (conditioned on independently sampled musical and code-style seeds) and the MIDI produced by rendering that program with the Strudel runtime. See Appendix A of the paper for the full construction pipeline.

Statistics

	Train	Test
# of (MIDI, Code) Pairs	20,152	1,022
# of Instruments	122	98
Avg. # Insts / Song	3.73±1.22	3.68±1.26
Avg. Lines of Code	52.20±20.88	52.46±21.70
Avg. Duration (s)	24.17±19.03	24.10±18.70

The ~20K training split is divided into two disjoint ~10K halves, one per post-training stage: SFT uses the first half as paired (MIDI, Strudel) examples, and RL uses the second half (MIDI only).

Contents

File	Description
dataset.jsonl	one record per pair: paths, split/subset assignment, tempo/meter metadata
strudel_synth.tar.gz	all pair files, organized in the three directories below

Extracting strudel_synth.tar.gz yields:

Directory	Description
midi/	*.mid — MIDI renders of each Strudel program
cycle_midi/	*.mid — copies of midi/ with onsets rescaled to cycle coordinates (paper Appendix B). Not for listening
answers/	*.js — the Strudel programs

dataset.jsonl fields

Field	Description
id	pair id (filenames share this id)
split	train / eval (eval is the paper's test split)
subset	sft / rl (train only): which post-training stage the pair belongs to
midi_path	path into the extracted midi/ directory
cycle_midi_path	path into the extracted cycle_midi/ directory
answer_path	path into the extracted answers/ directory
bpm, meter	tempo and beats-per-measure of the pair
original_midi_duration	duration of the MIDI in seconds

Usage

hf download anonymousgiraffe/Strudel-Synth --repo-type dataset --local-dir strudel-synth
cd strudel-synth
tar xzf strudel_synth.tar.gz

import json

pairs = [json.loads(l) for l in open("dataset.jsonl")]
train = [p for p in pairs if p["split"] == "train"]
test  = [p for p in pairs if p["split"] == "eval"]
print(len(train), len(test))  # 20152 1022

p = train[0]
midi_file = p["midi_path"]              # e.g. midi/000000.mid
strudel_program = open(p["answer_path"]).read()