LIBRA (Long Input Benchmark for Russian Analysis)
收藏资源简介:
LIBRA数据集由SaluteDevices等机构创建,旨在评估大型语言模型在俄语长文本理解方面的能力。该数据集包含21个子集,覆盖从4k到128k令牌的不同上下文长度,涉及多种复杂度和技能测试。数据集的创建过程包括翻译现有数据集、适应长输入任务以及基于开放数据创建新数据集。LIBRA数据集主要应用于自然语言处理领域,特别是大型语言模型的长上下文理解和处理能力的评估。
The LIBRA dataset was developed by institutions including SaluteDevices, with the core goal of evaluating the long-text comprehension capabilities of Large Language Models (LLMs) for Russian language materials. This dataset contains 21 subsets, covering diverse context lengths ranging from 4k to 128k Tokens, and incorporates test tasks with varying complexity and skill requirements. The dataset creation workflow encompasses three main stages: translating existing datasets, adapting to long-input tasks, and generating new datasets based on open data. The LIBRA dataset is primarily utilized in the Natural Language Processing (NLP) domain, particularly for evaluating the long-context understanding and processing performance of LLMs.

- 1Long Input Benchmark for Russian AnalysisSaluteDevices, Ecom.tech, MIPT, AIRI · 2024年



