Benchmark datasets are important in evaluating various methods for source code generation tasks. In this paper, we present AST2CVCode, a new benchmark dataset to enhance deep learning models for source code generation. AST2CVCode is prepared especially for the computer vision field. This dataset is used to develop and evaluate new methods for source code generation that can be applied to computer vision applications. To build the dataset from abstract syntax trees (ASTs), we develop a light-weight static construction tool to build the dataset. First, we collect Python codes for classification and recognition tasks. Second, the tool is used to convert these codes to their corresponding ASTs. Third, ASTs are interpreted to the dataset lines. Finally, the dataset is stored in a text-formatted file. We believe that the AST2CVCode dataset helps the developers to apply source code generation for real-world computer vision software development.