[ICLR 2026] [DCASE 2026 Training Set] AudioMCQ-StrongAC-GeminiCoT

This dataset is a highly curated subset of the AudioMCQ dataset, containing samples with native Chain-of-Thought (CoT) reasoning from Gemini 3.1 Pro that were answered correctly.

Note: The native CoT reasoning has been summarized by Gemini's internal algorithm before output, yet still retains rich audio details including timestamps, acoustic descriptions, and step-by-step temporal analysis.

🏆 DCASE 2026 Challenge: We are proud to announce that AudioMCQ-StrongAC-GeminiCoT serves as the official training set for DCASE 2026 Challenge Task 5, Audio-Dependent Question Answering.

News

• 2025-03-31: Added more data and filtered out low-quality CoT samples (e.g., hallucinations claiming no audio access or visual hallucinations). Now contains 19,480 samples.
• 2025-03-29: Initial release with 16,407 samples.

Dataset Origin

This dataset builds upon the foundation of our ICLR 2026 paper, "Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models".

1. Base Dataset: AudioMCQ's Strong Audio-Contribution (StrongAC) split.
   • Construction (Multi-model Consensus): Tested 3 LALMs using silent audio. If ≥2 models answered correctly, the sample was marked "weak" (audio not needed). The remaining audio-dependent samples form the StrongAC split.
2. Filtering Process:
   • Sampled data from the StrongAC split.
   • Used Gemini 3.1 Pro for audio question-answering inference.
   • Filtered samples where the model answered correctly and produced native chain-of-thought reasoning.

Statistics

Audio Source Distribution

Directory Structure

AudioMCQ-StrongAC-GeminiCoT/
├── data.jsonl              # Main data file
├── AudioCaps/              # Audio files
├── Clotho/
├── CompA-R/
├── LP-MusicCaps-MTT/
├── MusicCaps/
├── SpeechCraft/
└── Tacos/

Data Format

Each entry is a JSON object with the following fields:

Example

{
  "source_dataset": "Tacos",
  "id": "387817",
  "question_type": "temporal",
  "audio_path": "Tacos/387817.mp3",
  "question": "What other sound is present during the last engine sound?",
  "answer": "Water splashing",
  "choices": [
    "Nothing else can be heard",
    "Glass breaking",
    "Water splashing",
    "Wind blowing"
  ],
  "gemini_cot": "**Analyzing the Audio for the Concurrent Sound**\n\nOkay, so the challenge is to pinpoint the sound accompanying the final engine noise in this audio sequence. Let's break this down systematically. First, I need to analyze the sound file, and I've got a decent understanding from the descriptions, but I'll make sure to get all the data. I know there are repeated engine sounds mixed with a ticking sound. Let's note the time intervals.\n\n* From 00:00 to 00:02, it's just ticking. Then, 00:02 to 00:06, we get the engine *and* a water splashing sound.\n* Again, 00:06 to 00:10, more ticking. Then from 00:10 to 00:14, another burst of engine sound *and* that water splashing.\n* Then it is more ticking from 00:14 to 00:18, and finally, we get another engine sound from 00:18 to 00:23, where that water splashing is present again, according to the breakdown. And the last interval is a ticking sound from 00:23 to 00:26.\n\nThe question explicitly asks about the last engine sound. That's the one between 00:18 and 00:23. Listening closely to that part, I distinctly hear the sloshing sound, like water being agitated. It sounds like something mechanical is involved, which would make sense alongside the engine noises. A washing machine or something similar? Perhaps.\n\nNow, let's look at the given options to confirm. \"Nothing else can be heard\" is incorrect because I definitely hear the water. \"Glass breaking\" is clearly wrong – I don't hear any shattering or impact sounds. \"Water splashing\" is what I've identified, so that's the correct match. Finally, there's no wind, so \"Wind blowing\" is also incorrect.\n\nSo, the definitive answer is the water splashing sound."
}

Use Cases

This dataset is particularly valuable for:

• Audio Understanding Model Training: Leveraging Gemini's high-quality reasoning annotations as a strong supervisory signal.
• Chain-of-Thought Distillation: Transferring reasoning capabilities from large multimodal models to smaller, efficient models.
• Audio QA Benchmarking: Evaluating model performance on tasks requiring deep, step-by-step audio understanding.
• DCASE 2026 Task 5 Development: Serving as the foundational training data for participants.

License

This dataset follows the license agreement of the original AudioMCQ dataset, which is distributed under the Apache-2.0 license.

Citation

If you use AudioMCQ-StrongAC-GeminiCoT in your research or participate in the DCASE 2026 challenge using this dataset, please cite our original AudioMCQ paper:

@article{he2025measuring,
  title={Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models},
  author={He, Haolin and Du, Xingjian and Sun, Renhe and Dai, Zheqi and Xiao, Yujia and Yang, Mingru and Zhou, Jiayi and Li, Xiquan and Liu, Zhengxi and Liang, Zining and others},
  journal={arXiv preprint arXiv:2509.21060},
  year={2025}
}

@article{he2026summary,
  title={Summary of DCASE 2026 Task 5: Audio-Dependent Question Answering},
  author={He, Haolin and Sun, Renhe and Dai, Zheqi and Du, Xingjian and Wu, Chunyat and Liang, Zining and Liu, Zhengxi and Lei, Jiahe and Wang, Runbang and Zhou, Jiayi and Yang, Mingru and Li, Xiquan and Chen, Yun and Chen, Xie and Duan, Zhiyao and Wang, Weiqiang and Plumbley, Mark D. and Liu, Jian and Kong, Qiuqiang},
  journal={arXiv preprint arXiv:2607.18718},
  year={2026}
}