AIDC-AI/Marco-Bench-MIF

Marco-Bench-MIF是一个专为多语言指令遵循评估设计的深度本地化多语言基准，覆盖了30种语言，包括六大语系。该数据集通过词汇替换、主题转换和语用重构的三步流程，确保了文化和语言的适当性。它包含了541个指令-响应对，涵盖了单一/多个约束、表达/内容约束以及多种指令类型。数据集还提供了特定语言的机器翻译和文化定位版本，以便进行比较研究。

Marco-Bench-MIF is a deeply localized multilingual benchmark designed for instruction-following evaluation across 30 languages, spanning six major language families. It ensures cultural and linguistic appropriateness through a three-step process of lexical replacement, theme transformation, and pragmatic reconstruction. The dataset includes 541 instruction-response pairs covering single/multiple constraints, expressive/content constraints, and various instruction types, with machine-translated and culturally-localized versions available for comparative research.