M4LE

M4LE是一个多能力、多范围、多任务、多领域的长上下文评估基准，基于包含36个NLP数据集、11种任务类型和12个领域的多样化NLP任务池。为了缓解自然长序列任务的稀缺性并整合多能力评估，提出了一种自动方法（几乎无需人工标注）将短序列任务转换为统一的长序列场景，其中LLMs必须在基于显式或语义提示的长上下文中识别单个或多个相关跨度。

M4LE is a multi-capability, multi-scope, multi-task, and multi-domain long-context evaluation benchmark built on a diverse pool of NLP tasks encompassing 36 NLP datasets, 11 task types, and 12 domains. To alleviate the scarcity of natural long-sequence tasks and facilitate integrated multi-capability evaluation, an automatic method requiring nearly no manual annotation is proposed to convert short-sequence tasks into a unified long-sequence scenario, wherein LLMs must identify single or multiple relevant spans within long contexts grounded in explicit or semantic prompts.