Tevatron/docmatix-ir
收藏Hugging Face2024-08-12 更新2024-12-14 收录
下载链接：
https://hf-mirror.com/datasets/Tevatron/docmatix-ir
下载链接
链接失效反馈官方服务：
资源简介：
Docmatix-IR是一个适合训练文档视觉嵌入模型的数据集，用于开放域视觉问答任务。它通过对原始Docmatix数据集进行问题过滤和困难负例挖掘，生成了5.61M高质量的训练示例。原始Docmatix数据集包含2.4M PDF图像和9.5M相关问题，但许多问题不适合开放域问答。Docmatix-IR使用Document Screenshot Embedding (DSE)模型对语料库进行编码，并过滤掉不适合开放域检索的问题，同时挖掘困难负例以提高训练数据的质量。
Docmatix-IR is a dataset derived from the original Docmatix dataset, primarily designed for training document visual embedding models for open-domain visual question answering tasks. This dataset creates high-quality training data by filtering out unsuitable questions for open-domain question answering and mining hard negatives, ultimately obtaining 5.61M high-quality training examples.
提供机构：
Tevatron
