README.md

---
license: other
license_name: msp-podcast
license_link: https://lab-msp.com/MSP/MSP-Podcast.html
language:
- en
task_categories:
- audio-classification
tags:
- speech-emotion-recognition
- preference
- reasoning
- chain-of-thought
- dpo
pretty_name: Comparative Reasoning Traces for Preference-based SER
size_categories:
- 10K<n<100K
configs:
- config_name: arousal
  data_files: data/arousal.jsonl
- config_name: valence
  data_files: data/valence.jsonl
- config_name: dominance
  data_files: data/dominance.jsonl
---

# Comparative Reasoning Traces for Preference-based SER

Reasoning traces used to train the **SFT-CoT** and **DPO-CoT** models of
[Comparative Reasoning: Making an Audio Language Model Better at Comparing Emotions](https://arxiv.org/abs/2606.24082) (Interspeech 2026).

Models: [SFT](https://huggingface.co/Lab-MSP/comparative-reasoning-sft) · [SFT-CoT](https://huggingface.co/Lab-MSP/comparative-reasoning-sft-cot) · [DPO](https://huggingface.co/Lab-MSP/comparative-reasoning-dpo) · [DPO-CoT](https://huggingface.co/Lab-MSP/comparative-reasoning-dpo-cot)

Each example is an MSP-Podcast v2.0 training pair (10k pairs per attribute) with

- a **chosen** trace that explains why the clip with the higher attribute value is the answer, and
- a **rejected** trace that argues for the wrong clip (used as the rejected response in DPO-CoT).

The traces were generated by `Qwen/Qwen3-Next-80B-A3B-Thinking-FP8` from Qwen3-Omni-Captioner audio
captions and GeMAPS acoustic features (discretized into low / medium / high) of both clips, and
kept only if their final answer was consistent with the label.

The audio is **not** included. Obtain MSP-Podcast (release 2.0) from the
[MSP Lab](https://lab-msp.com/MSP/MSP-Podcast.html) under its license.

## Fields

| Field | Description |
|---|---|
| `key` | `<sen1 stem>_<sen2 stem>` |
| `emotion` | `arousal`, `valence`, or `dominance` |
| `sen1`, `sen2` | MSP-Podcast file names of Clip 1 and Clip 2 |
| `preference` | `1` if Clip 1 has the higher attribute value, `0` if Clip 2 does |
| `question` | Question asked to the audio LM, after the two clips (Clip 1 = `sen1`, Clip 2 = `sen2`) |
| `answer` | `Clip 1` or `Clip 2` (ground truth) |
| `reasoning` | Chosen reasoning trace |
| `reasoning_source` | `blind`: the reasoning LLM was not given the answer and reached the correct one; `guided`: regenerated with the ground-truth answer given, for pairs the blind mode got wrong |
| `rejected_answer` | The wrong answer |
| `rejected_reasoning` | Reasoning trace generated with the wrong answer given (`null` for 5 pairs where verification failed) |

## Training format

The audio LM gets a system prompt, the two clips, and `question`:

```
system: You are a helpful assistant for emotion comparative reasoning.
user:   <audio: sen1> <audio: sen2> {question}
```

The training response is `<think> {reasoning} </think> <answer> {answer} </answer>` (SFT-CoT, DPO-CoT chosen), and
`<think> {rejected_reasoning} </think> <answer> {rejected_answer} </answer>` for the DPO-CoT rejected response.

## Usage

```python
from datasets import load_dataset

traces = load_dataset("Lab-MSP/comparative-reasoning-traces", "arousal", split="train")
```

To build the ms-swift training data, download `data/*.jsonl` into a folder and run
`data/post_process/make_swift_data.py --traces_dir <folder>` from the code repository.

## Citation

```bibtex
@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}
}
```