temporal-lalm
收藏资源简介:
Temporal LALM (Relative Temporal MCQA) 是一个用于评估大型音频语言模型在相对时序推理能力上的多选题数据集。其核心任务是基于给定的音频片段,回答关于其中声音事件相对时序关系的选择题,具体包括判断哪个声音事件开始最早(earliest_start)、哪个结束最晚(latest_end)以及哪个持续时间最长(longest_duration)。该数据集构建于 TACOS 音频集合之上,共包含 1,657 道多选题,涵盖三个子任务,其中 earliest_start 任务有 528 题,latest_end 有 499 题,longest_duration 有 630 题。这些题目基于 718 个唯一的音频片段生成,每个音频片段可能被用于多个不同任务。数据集中每个任务的正确答案选项分布接近均匀,以平衡选项偏差。每条数据样本包含解码后的音频数据、原始音频文件名、问题文本、四个选项文本、正确答案对应的字母、正确答案的完整文本以及所属的任务标签。该数据集适用于音频分类、问答等任务,特别用于研究和评测模型在理解音频中事件时序关系方面的能力。
Temporal LALM (Relative Temporal MCQA) is a multiple-choice dataset developed to evaluate the relative temporal reasoning capabilities of large audio-language models. Its core task involves answering multiple-choice questions regarding the relative temporal relationships between sound events within a given audio clip, specifically including identifying which sound event has the earliest start time (earliest_start), the latest end time (latest_end), and the longest duration (longest_duration). This dataset is constructed based on the TACOS audio corpus, containing a total of 1,657 multiple-choice questions across three subtasks: 528 for the earliest_start task, 499 for the latest_end task, and 630 for the longest_duration task. These questions are generated from 718 unique audio clips, and each audio clip may be used for multiple distinct tasks. The distribution of correct answer options for each task in the dataset is nearly uniform to mitigate option bias. Each data sample includes decoded audio data, the original audio filename, the question text, four option texts, the letter corresponding to the correct answer, the full text of the correct answer, and the associated task tag. This dataset is applicable to tasks such as audio classification and question answering, and is specifically designed for researching and evaluating models' ability to understand temporal relationships between events in audio.
Temporal LALM (Relative Temporal MCQA) 数据集概述
该数据集是一个用于评估大型音频语言模型(LALMs)在音频片段中进行相对时间推理能力的多选题(MCQA)数据集。它基于 TACOS 音频集合构建。
-
任务与规模:包含三个核心任务,共计 1657 个多选题。
earliest_start(528题):判断哪个声音事件开始最早。latest_end(499题):判断哪个声音事件结束最晚。longest_duration(630题):判断哪个声音事件持续时间最长。
-
数据构成:
- 共有 718 个独特的音频片段(
audio),格式为解码的 mp3。 - 正确答案选项(A/B/C/D)的分布在每个任务中保持近平衡。
- 每个音频片段可能出现在多个任务中。
- 共有 718 个独特的音频片段(
-
数据字段:
audio: 解码的音频数据。filename: 源 mp3 文件名。question: 题目提示语。option_a至option_d: 四个候选答案文本。correct_answer: 正确选项的字母(A/B/C/D)。correct_text: 正确选项的文本内容。task: 所属任务类型。
-
许可证:知识共享 4.0 国际许可协议 (
cc-by-4.0)。 -
语言:英语(
en)。 -
相关资源:
- 项目主页:https://apoorva-ak.github.io/temporal-reasoning-lalms/
- 研究论文:https://arxiv.org/pdf/2606.17417
- 基础数据集 TACOS 论文:https://arxiv.org/abs/2505.07609




