MIR
收藏资源简介:
MIR数据集是一个专注于多图像交错推理的基准,包含22,257个图像文本交错问答对,涵盖12个不同的场景。该数据集旨在增强多模态大语言模型(MLLMs)在复杂多模态场景中的理解和表现,通过分析复杂的图像文本关系来实现。MIR数据集包括138,277张图像,平均每个实例包含六张图像。为了提高MLLMs的多图像交错数据理解能力,MIR数据集为每个实例配备了五个系统化和结构化的推理步骤:摘要、标题、文本到区域引用、区域到区域关系、结论。基于推理步骤,MIR数据集还提出了一种阶段式课程学习方法,该方法优先训练简单样本,并逐步引入更难的任务,从而引导MLLMs从“简单”到“困难”逐步掌握复杂的推理过程。
The MIR Dataset is a benchmark dedicated to interleaved multi-image reasoning, containing 22,257 image-text interleaved question-answer pairs spanning 12 distinct scenarios. This dataset is designed to enhance the understanding and performance of multimodal large language models (MLLMs) in complex multimodal scenarios through the analysis of intricate image-text relationships. The MIR Dataset comprises 138,277 images, with an average of six images per instance. To improve the multi-image interleaved data understanding capability of MLLMs, the MIR Dataset equips each instance with five systematic and structured reasoning steps: summary, captioning, text-to-region reference, region-to-region relationship, and conclusion. Based on these reasoning steps, the MIR Dataset also proposes a staged curriculum learning approach, which prioritizes training on simple samples and gradually introduces more challenging tasks, thereby guiding MLLMs to master complex reasoning processes progressively from "easy" to "hard".




