amazon-arabic-catalogue-v2
收藏资源简介:
该数据集包含阿拉伯语电子商务产品列表,配有教师生成的结构化JSON标签,旨在训练和评估输出符合模式的产品结构的小型语言模型(SLMs)。每个示例包括构建的`input_text`(阿拉伯语标题/描述及可选元数据)和`pred_json`标签,描述产品身份(品牌/类型/名称)、属性(颜色/材料)、包装(包装数量和净含量)、SEO关键词以及元置信度/警告。数据集适用于训练SLMs从阿拉伯语产品文本生成结构化电子商务JSON,评估结构化提取在保留验证和测试集上的表现,并通过`quality_tier_v2`和`warnings`进行课程学习或质量控制。标签由教师LLM使用模式引导的JSON生成(结构化输出)生成。数据分为训练集(24,000条)、验证集(3,008条)和测试集(3,008条)。
This dataset comprises Arabic e-commerce product listings paired with teacher-generated structured JSON labels, intended for training and evaluating small language models (SLMs) that produce structured product outputs aligned with predefined schemas. Each sample includes a constructed `input_text` (Arabic title/description with optional metadata) and a `pred_json` label that covers product identity (brand/type/name), attributes (color/material), packaging (packaging quantity and net content), SEO keywords, as well as meta confidence scores and warnings. The dataset is applicable for training SLMs to generate structured e-commerce JSON from Arabic product text, evaluating the performance of structured extraction on the held-out validation and test sets, as well as conducting curriculum learning or quality control using `quality_tier_v2` and `warnings`. The labels are generated by a teacher LLM via schema-guided JSON generation (structured output). The dataset is split into a training set (24,000 samples), a validation set (3,008 samples), and a test set (3,008 samples).




