license: apache-2.0
task_categories:
- video-text-to-text
- visual-question-answering
- question-answering
language:
- en
tags:
- omni
- audio
- video
- audio-visual
- audio-centric
pretty_name: AVUTBenchmark (AV-Human split, repackaged for lmms-eval)
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: test
path: av_human_data/test-*
dataset_info:
features:
- name: video_id
dtype: string
- name: question_id
dtype: string
- name: video_type
dtype: string
- name: task_type
dtype: string
- name: question
dtype: string
- name: options
dtype: string
- name: answer
dtype: string
- name: video_path
dtype: string
- name: subset_type
dtype: string
splits:
- name: test
num_bytes: 1050000
num_examples: 1734
AVUTBenchmark
This repository provides the AV-Human split of the AVUT benchmark in a format compatible with lmms-eval.
The data is provided as a single parquet file containing only the QA annotations. Since raw videos are not included, please download them from the original release and match them with the QA annotations using video_path.
Task configurations and evaluation scripts are available in the SEATS repository: https://github.com/xxayt/SEATS.
Overview
AVUT is an audio-centric video understanding benchmark that evaluates multimodal LLMs on both audio content comprehension and audio-visual interaction. It introduces an answer permutation-based filtering mechanism to eliminate the "text shortcut problem" where answers can be inferred from question text alone. We use the AV-Human partition containing expert-annotated QA pairs across 18 audio-centric domains.
| Statistic | Value |
|---|---|
| QA pairs | 1,734 (multiple choice, 4 options) |
| Videos | 691 videos, average ≈ 69 s |
| Video types | interview, music, education, sports, vlog, news, ... (18 domains) |
| Task types | Audio Character Matching, Audio OCR Matching, Audio Event Location, ... |
| Language | English |
| License | Apache-2.0 |
License
Released under the Apache-2.0 license. The upstream tsinghua-ee/AVUTBenchmark release does not specify a license, so this repackaged QA split defaults to Apache-2.0.