ViFoodLabel: A Vietnamese Food Product Label Dataset for Key Information Extraction

ViFoodLabel is a dataset of Vietnamese food product label images annotated for structured key information extraction. It comprises 600 images self-photographed with a smartphone camera (iPhone 13) from physical packaged-food products in Ho Chi Minh City, with no stock, e-commerce, or third-party packaging images included. Each image is paired with a hand-annotated ground-truth record covering nine fields: product name, ingredients, additives, warnings, a name/value nutrition table, origin, net weight, manufacturing date, and expiry date. A 200-image double-annotation pass, scored with the same matching procedure used for model evaluation, gives a mean macro field F1 of 0.873 (lenient) and a nutrition pairing accuracy of 0.999. Images were screened for readability and target content, then annotated by hand following a written guideline covering ingredient-versus-additive classification, warning identification, and language preference on multilingual labels. Annotations are released as one JSON file per image, each field a single string, a list of strings, or, for nutrition, a list of name/value pairs; empty fields reflect content absent from the photographed frame rather than annotation gaps. Images are released downscaled from capture resolution to keep the archive a manageable size while preserving label legibility. The deposit is organized as follows: images/ # the 600 label photographs (JPEG), each with a zero-padded # four-digit identifier from 0001 to 0600. labels/ # the 600 ground-truth records (JSON), one per image, sharing # the image identifier.ViFoodLabel是一款面向结构化关键信息提取任务的越南食品包装标签图像标注数据集。该数据集包含600张由智能手机（iPhone 13）实拍的实体包装食品标签图像，采集自越南胡志明市，未包含库存、电商或第三方包装图像。每幅图像均对应一份人工标注的基准真值（ground truth）记录，涵盖9个字段：产品名称、配料表、添加剂、警示信息、名值对形式营养表、产地、净含量、生产日期以及保质期。其中200张图像进行了双标注流程，采用与模型评估一致的匹配规则进行评分，最终得到宏平均字段F1值为0.873（宽松匹配模式），营养成分配对准确率为0.999。 所有图像均先经过可读性与目标内容筛选，随后依照涵盖配料与添加剂分类、警示信息识别及多语言标签语言偏好的书面标注指南完成人工标注。标注结果以单张图像对应一份JSON文件的形式发布，每个字段可为单个字符串、字符串列表，其中营养字段为名值对列表；空字段代表实拍标签中无对应内容，而非标注遗漏。图像已从原始拍摄分辨率下调分辨率，以在保证标签可读性的前提下控制归档文件体积。 数据集归档结构如下： images/ # 包含600张标签实拍图像（JPEG格式），图像采用四位补零编号，编号范围为0001至0600。 labels/ # 包含600份基准真值（ground truth）标注记录（JSON格式），与图像一一对应，共享相同的图像编号。