Spaces:

---

haiyewon

/

decomposer-demo

Like 2

Runtime error

Dear Hugging Face team,

We would like to apply for a community GPU grant for our Space.
Decomposer is a post-training framework for symbolic music decompilation, recovering executable, editable programs from symbolic music (MIDI-to-Strudel decompilation). Given a MIDI input, Decomposer generates a Strudel program that reproduces the input music when executed, while remaining readable and editable as code. It's trained with a two-stage SFT + RL (GRPO) pipeline on our synthetic dataset, Strudel-Synth (21K+ paired examples).

This is an academic project from Carnegie Mellon University (by Yewon Kim, Apurva Gandhi, David Chung, Graham Neubig, and Chris Donahue).

Project page: https://yewon-kim.com/decomposer/
Paper: https://arxiv.org/abs/2607.01849
Model: https://huggingface.co/midi-decomposer/Decomposer-Qwen3-8B

Since inference requires running an 8B model to generate and render Strudel code from MIDI in real time, a GPU-backed Space would let the community try live MIDI-to-code decompilation directly in the browser. Thank you so much for considering our request!