File size: 3,611 Bytes
11a7e82 b63adb3 11a7e82 b63adb3 11a7e82 b63adb3 11a7e82 b63adb3 11a7e82 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 | ---
license: other
license_name: usc-research
license_link: LICENSE
language:
- en
library_name: transformers
base_model: Qwen/Qwen2-Audio-7B-Instruct
tags:
- audio
- speech
- audio-llm
- paralinguistic
- pclm
- dpo
- voxparadox
pipeline_tag: audio-text-to-text
---
# Qwen2-Audio + PCLM + DPO
[](https://icml.cc/Conferences/2026)
[](https://arxiv.org/abs/2605.27772)
[](https://voxparadox.github.io/)
[](https://github.com/ihp-lab/VoxParadox)
[](https://huggingface.co/datasets/IHP-Lab/VoxParadox)
[](https://huggingface.co/IHP-Lab/AF3_PCLM_DPO)
[](LICENSE)
PCLM- and DPO-finetuned [Qwen2-Audio-7B-Instruct](https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct) from
*Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox*
(ICML 2026).
The base model is augmented with the **Prompt-Conditioned Layer Mixer (PCLM)** — a lightweight module that
adaptively mixes representations from intermediate audio-encoder layers based on the user prompt — and then
post-trained with **Direct Preference Optimization (DPO)** to prefer acoustically-grounded answers over
language-implied alternatives on paralinguistic MCQs.
## Usage
This checkpoint cannot be loaded with stock `transformers` — PCLM requires the custom
modeling code shipped in the [release repo](https://github.com/ihp-lab/VoxParadox).
```bash
git clone https://github.com/ihp-lab/VoxParadox
cd VoxParadox
conda create -n qwen2audio python=3.10 -y && conda activate qwen2audio
pip install torch torchaudio transformers accelerate librosa soundfile
```
Inference on VoxParadox (or any MCQ JSON in the same schema):
```bash
python -m qwen2audio.eval.run_eval \
--model_path IHP-Lab/Qwen2-Audio_PCLM_DPO \
--data_path /path/to/voxparadox.json \
--audio_base /path/to/audio_root \
--output_dir runs/eval/qwen2audio_pclm_dpo
```
Score with the dataset-shipped `eval.py`:
```bash
python eval.py --predictions runs/eval/qwen2audio_pclm_dpo/predictions.jsonl
```
The loader auto-detects `use_pclm=True` from `config.json` and activates PCLM with
`expose_layers=[5, 15, 25, 30]` over the audio encoder.
## Project resources
| Resource | Link |
|---|---|
| Paper (arXiv) | <https://arxiv.org/abs/2605.27772> |
| Project page | <https://voxparadox.github.io/> |
| Code | <https://github.com/ihp-lab/VoxParadox> |
| Benchmark | <https://huggingface.co/datasets/IHP-Lab/VoxParadox> |
| Sibling model (AF3) | <https://huggingface.co/IHP-Lab/AF3_PCLM_DPO> |
## Citation
```bibtex
@inproceedings{pang2026voxparadox,
title = {Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox},
author = {Pang, Jiacheng and Chaubey, Ashutosh and Soleymani, Mohammad},
booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
year = {2026}
}
```
## License
USC Research License (research / non-profit only). See [`LICENSE`](LICENSE).
The base model (`Qwen/Qwen2-Audio-7B-Instruct`) carries its own Tongyi Qianwen license terms,
which continue to apply to the inherited weights.
|