Karel DSL 和 Calculator DSL

本文介绍了两个用于神经程序合成的数据集：Karel DSL 和 Calculator DSL。Karel DSL 是一个包含复杂控制流和丰富输入空间的领域特定语言，用于训练深度网络以提高跨分布泛化性能。Calculator DSL 则是一个程序诱导任务，涉及计算简单算术表达式的结果。这两个数据集通过控制和评估合成数据分布的偏差，帮助改进深度网络的泛化能力。数据集的创建涉及定义随机变量来编码描述数据的特征，如 Karel 领域的控制流嵌套程度和输入中标记的数量。这些数据集的应用领域旨在解决自动生成特定语言程序的问题，通过训练深度网络在不同数据分布上提高性能。

This paper introduces two datasets for neural program synthesis: Karel DSL and Calculator DSL. Karel DSL is a domain-specific language (DSL) featuring complex control flows and a rich input space, designed to train deep networks for improved cross-distribution generalization performance. Calculator DSL, by contrast, is a program induction task that involves computing the results of simple arithmetic expressions. These two datasets help enhance the generalization capability of deep networks by controlling and evaluating the bias of their synthetic data distributions. The creation of these datasets entails defining random variables to encode the descriptive features of the data, such as the degree of control flow nesting in the Karel domain and the number of tokens in the inputs. The application scenarios of these datasets aim to address the problem of automatically generating programs in specific languages, by training deep networks to achieve improved performance across different data distributions.