Long-Range Arena
收藏资源简介:
Long-Range Arena是由谷歌研究创建的一个系统性统一基准,专注于评估模型在长上下文场景下的质量。该数据集包含从1K到16K令牌的序列,涵盖文本、自然、合成图像和数学表达式等多种数据类型和模态,要求模型进行相似性、结构和视觉空间推理。数据集创建过程中,设计了一系列具有特定内在结构的探测任务,以评估模型在不同类型数据和条件下的能力。Long-Range Arena旨在解决Transformer模型在处理长序列时的效率问题,为未来更高效架构的研究提供挑战和灵感。
Long-Range Arena is a systematic and unified benchmark developed by Google Research, dedicated to evaluating model performance in long-context scenarios. This benchmark includes sequences ranging from 1K to 16K Tokens, covering multiple data types and modalities such as text, natural images, synthetic images, and mathematical expressions, and requiring models to perform similarity, structural, and visual-spatial reasoning. During the creation of this benchmark, a series of probing tasks with specific inherent structures were designed to evaluate models' capabilities across different data types and conditions. Long-Range Arena aims to address the efficiency issues of Transformer models when processing long sequences, and provide challenges and inspirations for future research on more efficient architectures.




