README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-Omni-7B
tags:
  - lora
  - qwen2.5-omni
  - cantonese
  - speech
  - multimodal
---

# Qwen2.5-Omni-7B LoRA — mixed (Cantonese)

LoRA adapter fine-tuned on **multimodal_yue_benchmark** (Cantonese audio + text), speaker **all three (hiugaai/hiumaan/wanlung)**.

## Base model

Load with [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) as `model_name_or_path`, then load this repo as the PEFT adapter.

## Training

- Framework: [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory)
- Method: LoRA (r=8), bf16, DeepSpeed ZeRO-2
- Dataset: `wanlung_train` / `hiumaan_train` / `hiugaai_train` (single-speaker split)

## Inference (Transformers + PEFT)

```python
from transformers import AutoProcessor, Qwen2_5OmniForConditionalGeneration
from peft import PeftModel
import torch

base = "Qwen/Qwen2.5-Omni-7B"
adapter = "J017athan/Qwen2.5-Omni-7B-14k-mixed"

processor = AutoProcessor.from_pretrained(base, trust_remote_code=True)
model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    base, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(model, adapter)
```

Or use LLaMA-Factory `adapter_name_or_path` pointing to this repo.