MINT

该数据集是一个评估大型语言模型在多轮交互中使用工具和自然语言反馈解决问题的能力的基准。它包含了专注于推理、编程和决策制定的多样化评估数据集，这些数据集被精心挑选并整合成一个紧凑的子集，以便进行高效的评估。该任务的目的是评估大型语言模型的多轮交互能力。This dataset is a benchmark for evaluating the ability of large language models (LLMs) to solve problems through tool usage and natural language feedback during multi-turn interactions. It includes diverse evaluation datasets focused on reasoning, programming, and decision-making. These datasets have been carefully selected and consolidated into a compact subset to enable efficient assessment. The objective of this task is to evaluate the multi-turn interaction capabilities of large language models.

搜集汇总

背景与挑战

背景概述

MINT是一个多模态图像和叙事文本配乐数据集，专为配乐音频内容规划与生成而设计。它包含图像、叙事文本、音频标题和音频元数据，数据以JSON格式组织，图像来自Zenodo，音频从YouTube提取，持续时间为10秒。该数据集适用于学术研究，遵循CC BY-NC-SA-4.0许可，并强调伦理合规和版权保护。