---
title: MAJEPPA Conditional Generation
emoji: 🎹
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.13.0
app_file: app.py
pinned: false
license: cc-by-nc-4.0
short_description: Conditional piano performance generation with LLM-JEPA.
hardware: zero-a10g
sleep_time: 300
---

# MAJEPPA — Conditional Piano Performance Generation

Anonymized HF Space accompanying the MAJEPPA paper submission. Generates expressive
performance MIDIs from deadpan score MIDIs, conditioned on **performer type** and
**recording context**.

## How it works

- Base: `loubb/aria-medium` (660M params, MIDI tokenizer with bar/onset/duration tokens).
- Fine-tune: LoRA r=512 on `mixed_qkv` + `att_proj_linear`.
- Special tokens: `<COND_{performer}>`, `<COND_{recording}>`, `<PRED_1..k>`.
- Inference prompt: `[Score] [EOS] [COND_perf] [COND_rec] [PRED_1] [first_score_note]` →
  manual autoregressive sampling (temperature 0.8, top-k 50) until EOS.

## Hardware

GPU recommended (T4-small or better). Each generation runs ~512 manual decode
steps; on CPU this is ~30–60s per call.

## Files

```
app.py             # Gradio UI, lazy-loads the model on first request
inference.py       # Inference wrapper: load + generate(score, performer, recording)
src/models/        # Model code (slim): backbone + LoRA wrapper + sequence builder
examples/          # Five preset score MIDIs from the paper figures
requirements.txt
```