ALBA
收藏资源简介:
ALBA是由新里斯本大学语言专家团队开发的欧洲葡萄牙语(pt-PT)语言学评估基准,包含800条人工构建的测试问题,覆盖语言变体、文化语义、话语分析等8个语言学维度。该数据集通过专家标注的参考答案(720条)和LLM-as-a-judge框架支持生成式模型的细粒度评估,旨在解决pt-PT语言资源匮乏导致的模型偏见问题,推动葡萄牙语NLP工具的发展。数据来源于专家设计的原创任务,涉及方言转换、谚语解释、诗歌创作等典型应用场景。
ALBA is a European Portuguese (pt-PT) linguistic evaluation benchmark developed by a team of language experts from Universidade Nova de Lisboa. It contains 800 manually constructed test questions covering 8 linguistic dimensions including language variation, cultural semantics, discourse analysis and others. The dataset supports fine-grained evaluation of generative models via expert-annotated reference answers (720 in total) and the LLM-as-a-judge framework, aiming to address the model bias problem caused by the scarcity of pt-PT language resources and promote the development of Portuguese NLP tools. The data is sourced from original tasks designed by experts, involving typical application scenarios such as dialect conversion, proverb interpretation and poetry creation.



