Dyn-VQA
收藏资源简介:
Dyn-VQA数据集由阿里巴巴集团创建,旨在评估多模态检索增强生成(mRAG)系统在处理动态检索需求问题上的表现。该数据集包含1452个动态问题,涵盖快速变化答案、多模态知识和多跳问题三种类型。数据集的创建过程包括文本问题编写、多模态重写和中文-英文翻译三个步骤,确保了数据集的高质量和复杂性。Dyn-VQA主要应用于解决多模态大语言模型中的“幻觉”问题,通过提供复杂的知识检索策略来增强模型的适应性和准确性。
The Dyn-VQA dataset was created by Alibaba Group, aiming to evaluate the performance of multimodal retrieval-augmented generation (mRAG) systems when handling questions that require dynamic retrieval. This dataset contains 1452 dynamic questions, covering three categories: questions with rapidly changing answers, multimodal knowledge-based questions, and multi-hop questions. The construction of the dataset consists of three steps: text question compilation, multimodal rewriting, and Chinese-English translation, which ensures the high quality and complexity of the dataset. Dyn-VQA is primarily applied to address the "hallucination" problem in multimodal large language models, enhancing the adaptability and accuracy of the models by providing sophisticated knowledge retrieval strategies.




