ADQA-Bench: Audio-Dependent Question Answering Evaluation Set for DCASE 2026

ADQA-Bench is the official evaluation set for the DCASE 2026 Challenge Task 5: Audio-Dependent Question Answering. It focuses on addressing textual hallucination in Large Audio-Language Models by requiring models to answer questions based on audio perception rather than linguistic priors. The dataset was authored by Harland and last updated on May 29, 2026.

Use Cases

• Benchmarking model performance on audio-dependent question answering based on the described evaluation framework.
• Evaluating a model's susceptibility to textual hallucination based on the described task focus.
• Training or fine-tuning models for tasks requiring joint audio and language understanding based on the multimodal nature of the challenge.

Strengths

• Designed for a specific, rigorous evaluation framework targeting textual hallucination in LALMs.
• Serves as the official benchmark for a major international challenge (DCASE 2026 Task 5).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.

Provenance

Source

Harland on Hugging Face.

Collection Method

Created as the official evaluation set for DCASE 2026 Challenge Task 5.

Freshness

Last updated 2026-05-29 03:52:11; freshness should be verified.