Jianshu001/arabic-daily-batch01-cascade-5k

该数据集包含4348条阿拉伯语记录，经过多步骤处理，包括通过Gemma-4-31B进行级联重写、Gemma-as-rewriter清理、独立二进制判断和正则表达式后过滤。筛选标准包括避免系统提示回显、AI自我引用和用户对助手回答的简单附和。数据集的模式包括用户和助手的对话轮次、角色、思考和文本内容。

This dataset contains 4348 Arabic records, processed through multiple steps including cascade rewrite via Gemma-4-31B, Gemma-as-rewriter cleanup, independent binary judge, and final regex post-filter. The judge criteria include avoiding system-prompt echo, AI self-reference, and user sycophancy. The schema includes user and assistant turns, roles, thinking, and text.