---
license: mit
language:
- en
base_model:
- Qwen/Qwen2-Audio-7B-Instruct
---

- data process

- Place the WAV and JSON files in `dev_data`. 

To distinguish the recognition performance of each part, the file names of the training audio for Part One need to be prefixed with fold1-d-, those for Part Two need to be prefixed with fold1-a-, fold1-b-, fold1-c-, and those for Part Three need to be prefixed with fold1-e-. If the training audio file names for Part One and Part Three do not have the prefixes fold1-d- and fold1-e-, you will need to add them yourself. For example, if the file name of the training audio for Part One is 5402400A, then add the prefix to make it fold1-d-5402400A. If the file name of the audio for Part Three is audio_0001405, then add the prefix to make it fold1-e-audio_0001405. The process for the development set is the same, except that fold1 should be changed to fold2.

Download the pre-trained Sentence-BERT model and tokenizer from the following URL
and Place the downloaded pre-trained model and tokenizer inside the `../../qwen2_audio_baseline/Bert_pretrain`

- Example commands
  
```
git clone https://huggingface.co/PeacefulData/2025_DCASE_AudioQA_Baselines
cd 2025_DCASE_AudioQA_Baselines
mkdir Bert_pretrain
cd Bert_pretrain
git clone https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/tree/main
```


- Environment

```bash
cd ../qwen2_audio_baseline
pip install -r requirements.txt
```

You can also use a mirror source to speed up the process.
`pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple`

- Run Audio QA Inference Baseline
```bash
sh qwen_audio_test.sh
```