---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen3-8B
datasets:
  - haiyewon/Strudel-Synth
language:
  - en
tags:
  - music
  - midi
  - strudel
  - symbolic-music
  - music-decompilation
  - code-generation
  - qwen3
---

# Decomposer

**Decomposer** decompiles symbolic music (MIDI) into executable **Strudel** programs. 
This model is the **8B** model from the paper *Decomposer: Learning to Decompile Symbolic Music to Programs*, post-trained from **Qwen3-8B**.

- 📄 **Paper:** [arXiv:2607.01849](https://arxiv.org/abs/2607.01849)
- 🌐 **Project page:** [yewon-kim.com/decomposer](https://yewon-kim.com/decomposer)
- 🎹 **Live demo:** [haiyewon/decomposer-demo](https://huggingface.co/spaces/haiyewon/decomposer-demo)
- 💻 **Code:** [github.com/elianakim/Decomposer](https://github.com/elianakim/Decomposer)
- 🤗 **Dataset:** [haiyewon/Strudel-Synth](https://huggingface.co/datasets/haiyewon/Strudel-Synth)

## Model Description

- **Base model:** [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)
- **Model size:** 8.2B
- **Type:** Causal Language Model
- **Precision:** bfloat16

## Quick Start

Clone our [GitHub code repo](https://github.com/elianakim/Decomposer), run through the setup steps, and try:

```bash
git clone https://github.com/elianakim/Decomposer.git
cd Decomposer
uv sync

uv run generate_transformers.py \
    --model haiyewon/Decomposer-Qwen3-8B \
    --midi_path path/to/input.mid \
    --n_outputs 5
```

The repo and inference scripts provide a more complete usage guide. 
<!-- , including the
MIDI→text preprocessing (`MidiToText`) and the decompilation prompt -->


## Model Details
### Input Format

MIDI is serialized as instrument-wise note events. Onsets are given in
[**cycle coordinates**](https://strudel.cc/understand/cycles) instead of seconds:

```
[BPM=120 Meter=4]
[acoustic_grand_piano] C4@0.00 E4@0.25 G4@0.50 C5@1.00
[drum] bass_drum@0.00 closed_hi_hat@0.25 closed_hi_hat@0.50 snare@0.75
```

- **Header** `[BPM=… Meter=…]` — tempo/meter; if absent, [madmom](https://github.com/CPJKU/madmom) is used to estimate them.
- **Events** — one line per instrument, `pitch@cycle_onset` tokens; melodic pitches as note
  names (`C4`, `F#3`), drums as GM drum names (`bass_drum`, `closed_hi_hat`).


### Training Data

**[Strudel-Synth](https://huggingface.co/datasets/haiyewon/Strudel-Synth)** is a synthetic corpus of **21,174 (MIDI, Strudel) pairs**
(20,152 train / 1,022 test) built by distilling Strudel programs from Claude-Opus-4.6
and rendering each to MIDI with the Strudel runtime. Its ~20K training split is divided
into two disjoint ~10K halves, one per stage:

- **SFT** uses the first half as paired (MIDI, Strudel) examples.
- **RL** uses the second half (MIDI only), plus short-fragment (<30 s) [LMD](https://colinraffel.com/projects/lmd/) MIDI. 

### Inference Hyperparameters

Recommended settings matching the paper evaluation:

| Setting | Value |
|---|---|
| `temperature` | 1.0 |
| `max_new_tokens` | 4096 |
| `enable_thinking` | `false` |


## Citation

If you find our model useful, please cite our research as

```bibtex
@article{kim2026decomposer,
  title   = {Decomposer: Learning to Decompile Symbolic Music to Programs},
  author  = {Kim, Yewon and Gandhi, Apurva and Chung, David and Neubig, Graham and Donahue, Chris},
  journal = {arXiv preprint arXiv:2607.01849},
  year    = {2026}
}
```