TREA
收藏资源简介:
TREA数据集是一个针对音频的时序推理评估数据集,由印度科学学院电气工程系的研究人员创建。数据集包含600个音频样本,分为事件持续时间、事件排序和事件计数三个子任务,每个子任务包含200个样本。数据集来源于环境声音分类数据集ESC-50,每个录音时长为5秒。该数据集旨在评估大型音频语言模型在时序推理方面的能力,并揭示模型在理解音频事件时序信息方面的不足。
The TREA dataset is an audio-focused temporal reasoning evaluation dataset created by researchers from the Department of Electrical Engineering, Indian Institute of Science. It contains 600 audio samples, divided into three subtasks: event duration, event ordering, and event counting, with 200 samples per subtask. The dataset is derived from the environmental sound classification dataset ESC-50, and each audio recording has a duration of 5 seconds. This dataset aims to evaluate the temporal reasoning capabilities of large audio language models, and reveal the shortcomings of such models in understanding the temporal information of audio events.

- 1Benchmarking and Confidence Evaluation of LALMs For Temporal Reasoning印度科学学院电气工程系 · 2025年



