Reefknot
收藏资源简介:
Reefknot数据集由香港科技大学(广州)创建,专注于评估和缓解多模态大型语言模型中的关系幻觉问题。该数据集包含21,880个问题,跨越11,084张图像,来源于视觉基因组数据集,通过系统化的三元组识别和分类构建。数据集的创建过程包括三阶段的专家验证,确保数据质量。Reefknot数据集主要应用于提高多模态模型的可信度和准确性,特别是在处理复杂的关系推理任务时。
The Reefknot dataset was developed by The Hong Kong University of Science and Technology (Guangzhou), focusing on evaluating and mitigating relational hallucination issues in multimodal large language models. Comprising 21,880 questions and 11,084 images sourced from the Visual Genome dataset, it is constructed via systematic triplet recognition and classification. The dataset construction process includes three-stage expert validation to ensure data quality. The Reefknot dataset is primarily designed to enhance the credibility and accuracy of multimodal models, especially when tackling complex relational reasoning tasks.

- 1Reefknot: A Comprehensive Benchmark for Relation Hallucination Evaluation, Analysis and Mitigation in Multimodal Large Language Models香港科技大学(广州) · 2024年



