Jianshu001/arabic-daily-pipeline-v4-100

该数据集是一个阿拉伯语协议兼容的数据集，使用Gemma-4-31B模型进行自我评估。生成过程包括三个步骤：1) 使用v4协议兼容提示生成（包括助手和用户部分）；2) 通过Gemma作为重写器进行清理（移除草稿脚手架、角色标签等）；3) 对完整对话进行六维评估（真实性/助手质量/多轮/领域适应性/安全性/完整性），不符合要求的被丢弃。从100个生成的样本中保留了84个，丢弃了16个。丢弃原因包括真实性（13）、多轮（15）和领域适应性（1）。

This dataset is an Arabic protocol-compliant dataset, self-judged using the Gemma-4-31B model. The generation process includes three steps: 1) Generation using v4 protocol-compliant prompts (both assistant and user sides); 2) Cleanup via Gemma-as-rewriter (removes draft scaffolding, role labels, etc.); 3) 6-dim judge on full conversation (realism/assistant_quality/multi_turn/domain_fit/safety/integrity); dirty → drop. From 100 generated samples, 84 were kept and 16 were dropped. Breakdown of drops: realism=13, multi_turn=15, domain_fit=1.