GPTVQA
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/opendatalab/MLLM-DataEngine
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为GPTVQA，包含了问题-答案（Q-A）对，其中包括问题（Q）、多个选项（M）、答案（A）以及理由（E）。在实验中，探讨了两种格式：QMAE和QMA。该数据集是通过迭代精炼过程生成的，通过使用自适应坏案例采样和交互式提示优化，展现了提高的质量和多样性。数据集分两轮生成：第一轮生成5千条数据，第二轮生成1万8千条数据。该数据集的任务是视觉问答。
The dataset is named GPTVQA, which contains question-answer (Q-A) pairs including the question (Q), multiple-choice options (M), answer (A), and explanation (E). Two formats, QMAE and QMA, were explored in the experiments. This dataset is generated via an iterative refinement process, which adopts adaptive bad case sampling and interactive prompt optimization to achieve improved quality and diversity. It is generated in two rounds: 5,000 samples in the first round and 18,000 samples in the second round. The task of this dataset is visual question answering (VQA).
提供机构：
MLLM-DataEngine
