OpenLLM-France/EIFFEL
收藏资源简介:
EIFFEL(法语习语固定表达式大语言模型评估)是一个法语评估数据集,旨在评估大语言模型在语境中对法语习语表达的理解能力。该数据集包含602个多项选择题形式的样本,用于测试LLM在上下文中完成习语表达的能力。样本由母语者和语言学家手动构建和标注。数据集根据法语习语在英语中是否有(直接)翻译进行分类,分为三个类别:逐字翻译(88个表达)、相似翻译(100个表达)和不同翻译(414个表达)。每个样本包含多个字段,如带上下文的法语句子、法语习语、英语翻译(如可能)、掩码句子、四个答案选项、正确答案、类型、语域和数据来源。数据集旨在研究以英语为中心的培训对LLM在非英语语言(特别是涉及文化特定任务的内容)上的性能影响。
EIFFEL (Evaluation of Idiomatic French Fixed Expressions for Large Language Models) is a French evaluation dataset designed to assess large language models’ knowledge of idiomatic expressions in context. EIFFEL comprises 602 samples in multiple choice format designed to test an LLMs capacity to complete an idiomatic expression in context. The samples have been manually constructed and annotated by native speakers and linguists. The dataset is organized based on whether a given French idiomatic expression has a (direct) translation in English, with categories: word-for-word (88 expressions), similar (100 expressions), and different (414 expressions). Each sample includes fields such as French with context, French idiomatic expression, English translation, masked sentences, answer options, correct answer, type, register, and data origin. The dataset contributes to research on the impact of anglocentric training on LLM performance in non-English languages, especially for culturally-specific tasks.




