metadata
license: other
license_name: qwen-research
license_link: https://huggingface.co/Qwen/Qwen2.5-Omni-3B/blob/main/LICENSE
base_model: Qwen/Qwen2.5-Omni-3B
library_name: peft
tags:
- speech-emotion-recognition
- preference
- audio
- lora
Comparative Reasoning — SFT-CoT (Qwen2.5-Omni-3B LoRA)
The SFT-CoT model from Comparative Reasoning: Making an Audio Language Model Better at Comparing Emotions (Interspeech 2026). Given two utterances, it answers which one has higher arousal, valence, or dominance and explains why.
This repository contains a LoRA adapter (rank 64, alpha 128, all linear layers) for Qwen/Qwen2.5-Omni-3B.
| Training | SFT, lr 1e-4, total batch size 32 |
| Training data | 10k MSP-Podcast v2.0 preference pairs per attribute (30k total), target: reasoning trace + answer |
| MSP-Podcast test (A / V / D / Avg) | 85.5 / 86.5 / 84.6 / 85.5 |
| BIIC-Podcast / WHiSER (Avg) | 74.1 / 85.4 |
Usage
Merge the adapter and evaluate with the code repository:
swift export --adapters Lab-MSP/comparative-reasoning-sft-cot --merge_lora true \
--model Qwen/Qwen2.5-Omni-3B --output_dir outputs/experiments/sft-cot_merged
bash src/eval.sh msp_test outputs/experiments/sft-cot_merged
Prompt format (two audios followed by the question):
system: You are a helpful assistant for emotion comparative reasoning.
user: <audio><audio>You will hear two audio clips. Clip 1 is the first audio clip. Clip 2 is the second audio clip. <attribute definition> Which clip has higher <attribute>?
The model answers <think> ... </think> <answer> Clip 1 </answer> or ... Clip 2 ....
Citation
@inproceedings{naini2026comparative,
title = {Comparative Reasoning: Making an Audio Language Model Better at Comparing Emotions},
author = {Naini, Abinay Reddy and Kim, Jaeyeon and Yang, Chao-Han Huck and Watanabe, Shinji and Busso, Carlos},
booktitle = {Interspeech},
year = {2026}
}