Datasets:

gwkrsrch
/

music_avqa_hard

source_index int64 1 9.18k	question_id stringlengths 2 5	gpt4o_pred stringclasses 318 values	judge_pred stringclasses 1 value
1	13	Unknown	no
3	35	Two violins.	no
5	55	No.	no
6	66	I can't determine the number of instruments sounding from the image.	no
8	82	Two.	no
9	83	Three instruments.	no
11	89	No.	no
12	90	Two.	no
14	95	No.	no
16	103	No.	no
21	118	Unknown.	no
28	155	One.	no
30	160	No.	no
32	169	No.	no
34	184	Two.	no
35	188	No.	no
40	201	One.	no
43	224	In the image.	no
45	227	Three.	no
47	239	Two.	no
50	260	Accordion.	no
52	267	Four.	no
53	275	Three flutes.	no
55	281	Three.	no
57	288	One.	no
58	290	Two.	no
59	295	Three.	no
60	302	Two.	no
61	312	Multiple.	no
62	313	No.	no
65	324	No.	no
71	347	Two.	no
75	383	Four.	no
76	385	Four.	no
84	413	Three.	no
96	499	Three types.	no
97	500	Two.	no
99	523	Three.	no
100	528	One instrument.	no
102	537	Two ukuleles.	no
114	579	No.	no
123	615	Three.	no
124	618	No.	no
125	635	Two.	no
129	644	Two.	no
133	660	I can't determine the number of instruments sounding from the image alone.	no
142	723	Three.	no
144	744	One.	no
146	755	No.	no
156	786	Two.	no
158	790	No.	no
161	809	No.	no
162	810	No.	no
163	813	Three.	no
165	818	One.	no
166	829	I don't know.	no
168	849	Three.	no
171	856	Two.	no
181	911	No.	no
184	927	Accordion	no
185	937	No.	no
190	951	No.	no
194	977	No.	no
195	981	One.	no
196	983	No.	no
204	1026	No.	no
205	1036	Unknown	no
210	1052	No.	no
211	1057	Two instruments.	no
213	1069	Violin.	no
215	1079	No cello.	no
219	1099	Unknown.	no
221	1113	Violin.	no
222	1117	No.	no
223	1119	No.	no
228	1132	No.	no
229	1138	Three.	no
230	1138	Three.	no
231	1150	Five.	no
236	1163	No.	no
245	1206	No.	no
246	1210	Eight.	no
248	1219	One.	no
251	1227	Seven.	no
253	1233	One.	no
256	1250	No.	no
257	1257	No.	no
259	1272	No.	no
261	1280	No.	no
270	1354	One violin.	no
271	1365	One.	no
272	1367	Two.	no
279	1392	No.	no
284	1413	Two.	no
286	1417	No.	no
295	1453	Three.	no
296	1462	Unknown.	no
297	1464	Violin.	no
298	1470	No.	no
306	1534	Unknown.	no

End of preview. Expand in Data Studio

Music-AVQA-Hard

Music-AVQA-Hard: the 2,514 Music-AVQA test items whose question type involves audio (the official type does not contain "Visual") and whose single-frame GPT-4o answer the gpt-4o-mini judge rejected (judge_pred == "no", judge_score == 0).

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 Music-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/json/avqa-test.json of GeWu-Lab/MUSIC-AVQA), checks its SHA-256 and joins it onto these rows (UniMambaMia/official_annotations.py). Videos have to be obtained from the Music-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
`source_index`	0-based position of the item in the official `avqa-test.json` (the join key; `question_id` alone is not unique in that file).
`question_id`	`question_id` of that item, kept for reference.
`gpt4o_pred`	GPT-4o (`gpt-4o-2024-08-06`, temperature 0) answer when shown only the central frame, without audio.
`judge_pred`, `judge_score`	Verdict of the gpt-4o-mini judge (ActivityNet-QA prompt of lmms-eval) on that answer.

Usage

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

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

To use it outside lmms-eval:

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

ds = attach_music_avqa(datasets.load_dataset("gwkrsrch/music_avqa_hard", split="test"))

License

The columns of this dataset are released under the MIT License. Music-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 Music-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{li2022learning,
  title     = {Learning to Answer Questions in Dynamic Audio-Visual Scenarios},
  author    = {Li, Guangyao and Wei, Yake and Tian, Yapeng and Xu, Chenliang and Wen, Ji-Rong and Hu, Di},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2022}
}

Downloads last month: 46

Paper for gwkrsrch/music_avqa_hard

Does Audio Matter for Modern Video-LLMs and Their Benchmarks?

Paper • 2509.17901 • Published Sep 22, 2025