humain-ai/LC-Eval

LC-Eval是一个全面的双语多任务评估基准，旨在严格评估大型语言模型在长文本理解方面的能力，支持阿拉伯语和英语。该数据集包含从4K到超过128K令牌的上下文长度，旨在解决现有基准主要关注英语或依赖于合成任务的问题，并提供真实世界的挑战，需要深度推理、跨文档合成和双语理解。

LC-Eval is a comprehensive bilingual multi-task evaluation benchmark designed to rigorously assess the capabilities of Large Language Models (LLMs) in long-context understanding for both Arabic and English. The dataset includes context lengths from 4K to over 128K tokens, addressing the critical gaps in existing benchmarks that focus primarily on English or rely on synthetic tasks, and presents real-world challenges that require deep reasoning, cross-document synthesis, and bilingual comprehension.