lightonai/llamaindex-vdr-fine-tuning
收藏资源简介:
LlamaIndex VDR Fine-Tuning数据集是一个多语言视觉文档检索数据集,专为使用PyLate进行检索模型微调而设计。它基于llamaindex/vdr-multilingual-train数据集重新格式化,包含德语、英语、西班牙语、法语和意大利语的查询文本、文档元数据以及硬负例。数据集结构分为三个部分:queries表存储查询ID和查询文本;documents表存储文档ID和对应的图像文件名(图像存储在独立的lightonai/llamaindex-vdr-images数据集中);scores表存储每个查询的候选文档ID列表和相关性分数列表,其中第一个文档为正例(分数为1.0),其余为挖掘的硬负例(分数为0.0)。每种语言都有训练和测试分割,训练分割包含硬负例,测试分割仅含正例。硬负例使用voyage-3模型以0.75的相似度阈值挖掘。数据集适用于多模态检索任务,需结合图像数据集使用。
The LlamaIndex VDR Fine-Tuning dataset is a multilingual visual document retrieval dataset designed for retrieval fine-tuning with PyLate. It reformats the llamaindex/vdr-multilingual-train dataset, containing queries in German, English, Spanish, French, and Italian, along with document metadata and hard negatives. The dataset structure consists of three components: the queries table with query IDs and query texts; the documents table with document IDs and corresponding image filenames (images are stored separately in the lightonai/llamaindex-vdr-images dataset); and the scores table with lists of candidate document IDs and relevance scores for each query, where the first document is the positive (score 1.0) and the remaining are mined hard negatives (score 0.0). Each language has train and test splits, with train splits including hard negatives and test splits containing only positives. Hard negatives are mined using voyage-3 with a similarity threshold of 0.75. The dataset is suited for multimodal retrieval tasks and requires integration with the image dataset.




