Reformulation Training Set
收藏资源简介:
该数据集名为Reformulation Training Set,由耶路撒冷希伯来大学和墨尔本大学的研究团队创建,旨在通过人类反馈改进图像描述生成模型。数据集包含1405条数据,来源于MSCOCO和Flickr30k的测试集,数据生成过程涉及使用BLIP、mPLUG和ClipCap等模型生成初始描述,并由人类注释者进行修正。数据集的应用领域主要集中在图像描述生成,特别是非英语描述和风格转换任务,旨在通过反馈模型提升描述的事实准确性和风格适应性。
This dataset, named Reformulation Training Set, was developed by a research team from the Hebrew University of Jerusalem and the University of Melbourne, with the aim of improving image captioning models through human feedback. It consists of 1405 data entries sourced from the test sets of MSCOCO and Flickr30k. The data generation process entails generating initial captions using models including BLIP, mPLUG and ClipCap, followed by revisions made by human annotators. The dataset is mainly applied in the field of image captioning, especially non-English captioning and style transfer tasks, with the objective of boosting the factual accuracy and style adaptability of image descriptions via feedback models.




