FineCops-Ref
收藏资源简介:
FineCops-Ref数据集是由电子科技大学研究团队创建的,针对多模态大型语言模型(MLLMs)设计的参照表达式理解(REC)基准数据集。该数据集包含正样本和负样本,正样本由图像和文本表达式组成,负样本包含未配对的图像和文本表达式,用于评估模型在目标物体缺失的情况下的拒绝能力。数据集分为三个难度级别,要求MLLMs在不同属性和关系上进行多级细粒度推理。通过精细编辑和增强生成的负样本,进一步评估模型对齐错误和虚构情况下的鲁棒性。
The FineCops-Ref dataset is a referring expression comprehension (REC) benchmark designed for multimodal large language models (MLLMs), developed by a research team from the University of Electronic Science and Technology of China. It comprises positive and negative samples: positive samples consist of paired images and textual expressions, while negative samples consist of unpaired images and textual expressions, aimed at evaluating the model's rejection ability when the target object is absent. The dataset is categorized into three difficulty levels, which require MLLMs to conduct multi-level fine-grained reasoning across various attributes and relational cues. Through meticulously edited and augmented negative samples, the robustness of models against alignment errors and hallucinations is further assessed.

- 1New Dataset and Methods for Fine-Grained Compositional Referring Expression Comprehension via Specialist-MLLM Collaboration电子科技大学 · 2025年



