gwkrsrch/avqa_2025

AVQA (retrievable validation items)

The 9,167 items of the AVQA validation set (Yang et al., ACM MM 2022) whose clips could still be retrieved in 2025, as used in the paper below. AVQA-Hard (gwkrsrch/avqa_hard) is the subset GPT-4o answers incorrectly from the central frame.

Released with Do Modern Video-LLMs Need to Listen? A Benchmark Audit and Scalable Remedy (Interspeech 2026, arXiv:2509.17901); code: naver-ai/unimambamia-av.

What is (and is not) in this dataset

This dataset contains item identifiers and values we produced (single-frame probe outputs and the like). It does not contain any question, option or answer text of AVQA, nor any video. Those remain with the original release and its terms: the evaluation code of UniMambaMia-AV downloads the official annotation file at a fixed commit (data/annotation/val_qa.json of AlyssaYoung/AVQA), checks its SHA-256 and joins it onto these rows (UniMambaMia/official_annotations.py). Videos have to be obtained from the AVQA project.

Row order is fixed, so the doc_ids in benchmark_audit/single_frame_filter.json of the repository refer to the rows of this dataset in order.

Columns

Column	Description
id	Item id in the official AVQA val_qa.json.
data_source	Always avqa_val.
video_duration	Duration (s) of the clip we used, measured with ffprobe.
gpt4o_raw	Raw GPT-4o (gpt-4o-2024-08-06, temperature 0) response when shown only the central frame of the clip, without audio.
gpt4o_pred_letter, gpt4o_pred_index	Option parsed from that response.
gpt4o_correct	Whether that option is the correct one.

Usage

The lmms-eval task avqa_2025 of the repository loads this dataset and joins the official annotations automatically:

bash eval_unimambamia.sh gwkrsrch/UniMambaMia-AV-Qwen2-7B --tasks avqa_2025

To use it outside lmms-eval:

import datasets
from official_annotations import attach_avqa  # UniMambaMia/official_annotations.py

ds = attach_avqa(datasets.load_dataset("gwkrsrch/avqa_2025", split="test"))

License

The columns of this dataset are released under the MIT License. AVQA itself (annotations and videos) is not part of this dataset and is subject to the terms of its original release.

Citation

If you use this dataset, please cite our paper and the original AVQA paper.

@inproceedings{kim2026doesaudiomatter,
  title     = {Do Modern Video-LLMs Need to Listen? A Benchmark Audit and Scalable Remedy},
  author    = {Geewook Kim and Minjoon Seo},
  booktitle = {Proc. Interspeech 2026},
  year      = {2026},
  url       = {https://arxiv.org/abs/2509.17901}
}

@inproceedings{yang2022avqa,
  title     = {AVQA: A Dataset for Audio-Visual Question Answering on Videos},
  author    = {Yang, Pinci and Wang, Xin and Duan, Xuguang and Chen, Hong and Hou, Runze and Jin, Cong and Zhu, Wenwu},
  booktitle = {Proceedings of the 30th ACM International Conference on Multimedia},
  year      = {2022}
}