Simple reversals, Simple syllogisms, Reversal curse paper, Semantic structure benchmark
收藏资源简介:
本文提出了一系列数据集,旨在研究语言模型在上下文学习和微调中的泛化能力。这些数据集包括简单逆转、简单三段论、逆转诅咒论文和语义结构基准。它们被设计为隔离数据集中的知识,以创建清洁的泛化测试。这些数据集通过让预训练的大型模型接触数据集中的控制子集信息,并在需要各种泛化类型的测试集上评估其性能。研究结果表明,在数据匹配的情况下,上下文学习比微调更加灵活。此外,本文还提出了一种通过添加上下文推理来改进微调泛化的方法。
This paper presents a suite of datasets designed to investigate the generalization capabilities of language models in in-context learning and fine-tuning. These datasets cover Simple Reversal, Simple Syllogism, the Reverse Curse paper, and the Semantic Structure Benchmark. They are developed to isolate the knowledge within the datasets, thus creating clean, uncontaminated generalization tests. These datasets operate by exposing pre-trained large language models to information from their control subsets, and evaluating the models' performance on test sets that require diverse types of generalization. The research findings demonstrate that in data-matched settings, in-context learning is more flexible than fine-tuning. Furthermore, this paper proposes a method to enhance the generalization performance of fine-tuning by integrating contextual reasoning.




