README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: eval
        path: eval-noanswer.jsonl
license: apache-2.0

ADQA-Bench: Audio-Dependent Question Answering Evaluation Benchmark

This is the official Evaluation Set for DCASE 2026 Challenge Task 5: Audio-Dependent Question Answering (ADQA).

The ADQA task focuses on addressing "Textual Hallucination" in Large Audio-Language Models (LALMs) — where models pass audio understanding benchmarks by relying on text prompts and internal linguistic priors rather than actual audio perception. ADQA introduces a rigorous evaluation framework using Audio-Dependency Filtering (ADF) to ensure questions cannot be answered through common sense or text-only reasoning.

Audio-Dependency Filtering (ADF)

All samples in this evaluation set undergo a rigorous four-step ADF hard-filtering process to guarantee genuine audio dependence:

1. Silent Audio Filtering: Questions solvable by LALMs without audio are removed.
2. LLM Common-sense Check: Ensures no external knowledge alone can solve the question.
3. Perplexity-based Soft Filtering: Eliminates samples with text-based statistical shortcuts.
4. Manual Verification: Final human-in-the-loop check for ground-truth accuracy.

Statistics

Metric	Count
Total Samples	3,000
Unique Audio Files	3,000

Data Sources

The evaluation set is composed of two parts:

• Existing Benchmarks: A portion of the samples is derived from established audio understanding benchmarks, including MMAU, MMAR, and MMSU. These samples cover a wide range of audio understanding tasks such as speech, music, and sound perception.
• Human-Annotated Questions: The remaining portion consists of newly constructed, human-annotated multiple-choice questions based on diverse audio sources, designed to further challenge models on real-world audio comprehension.

All samples undergo the four-step Audio-Dependency Filtering (ADF) process described above.

Directory Structure

ADQA-Bench/
├── eval-noanswer.jsonl       # Evaluation data without answers (3,000 samples)
├── eval.jsonl                # Full data with answers (to be released after the competition)
├── eval_audios/              # Audio files (3,000 .wav files)
└── README.md

Data Format

Each entry in eval-noanswer.jsonl is a JSON object with the following fields:

Field	Type	Description
id	string	Unique sample identifier (e.g., eval_0001)
audio_path	string	Relative path to audio file
question_text	string	Question text
multi_choice	list[string]	Answer choices

Example

{
  "id": "eval_0001",
  "audio_path": "eval_audios/eval_0001.wav",
  "question_text": "What is the speaker's primary emotion in this audio?",
  "multi_choice": ["Sadness", "Happiness", "Anger", "Fear"]
}

License

This dataset is distributed under the Apache-2.0 license.

Citation

If you use this evaluation set or participate in DCASE 2026 Task 5, please cite:

@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}
}