SoundMind Dataset

SoundMind: RL-Incentivized Logic Reasoning for Audio-Language Models (EMNLP 2025)

SoundMind Dataset is an Audio Logical Reasoning (ALR) dataset consisting of 6,446 text-audio annotated samples specifically designed for complex reasoning tasks, where models determine whether conclusions are "entailed" or "not-entailed" based on logical premises, featuring comprehensive chain-of-thought reasoning annotations in both audio and text formats as the first audio-level reasoning dataset for training sophisticated audio-language models.

Github Paper Dataset Model

Dataset Summary

• Task: Logical reasoning and entailment classification
• Modalities: Text (annotations) + Audio (questions and answers)
• Language: English
• Total Samples: 6,446
• Splits: Train (5,184), Test (656), Dev (606)
• Labels: Binary (0: not-entailed, 1: entailed)

Dataset Structure

Data Splits

Directory Structure

dataset/
├── train/
│ ├── metadata.jsonl # Training metadata
│ ├── 0001/
│ │ ├── annotation.json # Complete annotation data
│ │ ├── question.wav # Audio question
│ │ └── answer_{0|1}.wav # Audio answer (based on label)
│ ├── 0002/
│ └── ... (5,184 samples)
├── test/
│ ├── metadata.jsonl # Test metadata
│ └── ... (656 samples)
└── dev/
  ├── metadata.jsonl # Validation metadata
  └── ... (606 samples)

Data Fields

metadata.jsonl Format

Each line in the metadata files contains:

{
  "id": "0001",
  "question_file_name": "0001/question.wav",
  "answer_file_name": "0001/answer_1.wav",
  "annotation_question": "Text description of the logical reasoning task...",
  "annotation_answer": "Detailed reasoning explanation...",
  "label": 1
  }

annotation.json Format

Each sample directory contains a complete annotation file:

{
  "user_content": "Logical reasoning task description...",
  "answer": "Step-by-step reasoning explanation...",
  "label": 1,
  "id": "sequential_identifier"
}

Detailed Case:

Input Question (Audio/Text): Prompt + three logical components:

• Major premise: General rule/definition (e.g., "harmful behavior of omission refers to...")
• Minor premise: Specific question/context (e.g., "which of the following acts is...")
• Conclusion: Specific scenario to evaluate (e.g., "D abandoned the baby...")

Output Answer (Audio/Text): Response with:

• Chain-of-thought: Step-by-step logical analysis
• Final answer: "entailed" or "not-entailed"

Field Descriptions

• id: Unique identifier for each sample (e.g., "0001", "0002")
• question_file_name: Relative path to the audio question file
• answer_file_name: Relative path to the audio answer file (answer_0.wav or answer_1.wav based on label)
• annotation_question: Text content of the logical reasoning question
• annotation_answer: Detailed step-by-step reasoning explanation
• label: Binary classification label (0: not-entailed, 1: entailed)

Citation

If you find our work helpful, feel free to give us a cite.

@article{diao2025soundmind,
  title={SoundMind: RL-Incentivized Logic Reasoning for Audio-Language Models},
  author={Diao, Xingjian and Zhang, Chunhui and Kong, Keyi and Wu, Weiyi and Ma, Chiyu and Ouyang, Zhongyu and Qing, Peijun and Vosoughi, Soroush and Gui, Jiang},
  journal={arXiv preprint arXiv:2506.12935},
  year={2025}
}