---
license: mit
language: en
tags:
- llm
- music
- multimodal
- midi
- phi-3
- question-answering
- optical-music-recognition
model-index:
- name: Phi-3-MusiX
  results: []
datasets:
- puar-playground/MusiXQA
pipeline_tag: image-text-to-text
base_model:
- microsoft/Phi-3-vision-128k-instruct
library_name: peft
---

# Phi-3-MusiX 🎵

**Phi-3-MusiX** is a LoRA adapter for [microsoft/Phi-3-vision-128k-instruct](https://huggingface.co/microsoft/Phi-3-vision-128k-instruct) for understanding symbolic music in the form of scanned music sheets, MIDI files, and structured annotations.
This adapter equips Phi-3 with the ability to perform symbolic music reasoning and answer questions about scanned music sheets and MIDI content.
- Sorce code: [GitHub](https://github.com/puar-playground/MusiXQA)
- Dataset: [MusiXQA](https://huggingface.co/datasets/puar-playground/MusiXQA)
- Paper: [arXiv](https://arxiv.org/abs/2506.23009)

---

## Inference
```
from transformers import AutoModelForCausalLM 
from transformers import AutoProcessor
from PIL import Image 
from http import HTTPStatus
import torch
import requests
from io import BytesIO

def load_img(img_dir):
  if img_dir.startswith('http://') or img_dir.startswith('https://'):
      response = requests.get(img_dir)
      image = Image.open(BytesIO(response.content)).convert('RGB')
  else:
      image = Image.open(img_dir).convert('RGB')
  return image


model = AutoModelForCausalLM.from_pretrained('microsoft/Phi-3-vision-128k-instruct', device_map="cuda", trust_remote_code=True, torch_dtype="auto")
processor = AutoProcessor.from_pretrained('microsoft/Phi-3-vision-128k-instruct', trust_remote_code=True)
model.load_adapter('puar-playground/Phi-3-MusiX')


prompt = '' + f'USER: Answer the question:\n{question_string}. ASSISTANT:'

# setup message
messages = [{"role": "user", "content": f"<|image_1|>\n{prompt}"}]

# load image from dir
image = load_img(img_dir)

prompt_in = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(prompt_in, [image], return_tensors="pt").to("cuda")

generation_args = { 
    "max_new_tokens": 500,
    "temperature": 0.1,
    "do_sample": False,
}

with torch.no_grad():
    generate_ids = model.generate(**inputs, eos_token_id=processor.tokenizer.eos_token_id, **generation_args)

# remove input tokens
generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
model_answer = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        
```



## 🧪 Training Data

The model is trained on the [MusiXQA](https://huggingface.co/datasets/puar-playground/MusiXQA) dataset, which includes four QA sets:

Each entry in the dataset includes:
- A scanned music sheet image
- Its structured metadata (`metadata.json`)
- A MIDI file
- QA pair targeting music understanding

---

## 🎓 Reference
If you use this dataset in your work, please cite it using the following reference:
```
@misc{chen2025musixqaadvancingvisualmusic,
      title={MusiXQA: Advancing Visual Music Understanding in Multimodal Large Language Models}, 
      author={Jian Chen and Wenye Ma and Penghang Liu and Wei Wang and Tengwei Song and Ming Li and Chenguang Wang and Jiayu Qin and Ruiyi Zhang and Changyou Chen},
      year={2025},
      eprint={2506.23009},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.23009}, 
}
```