CodeXGLUE is a comprehensive benchmark dataset designed to foster machine learning research in program understanding and generation. It includes 14 datasets and 10 tasks across various programming languages, covering clone detection, defect detection, cloze test, code completion, code repair, code-to-code translation, text-code, code-text, and text-text tasks. The dataset supports both supervised and unsupervised learning, making it suitable for a wide range of applications. CodeXGLUE also provides three baseline systems—BERT-style, GPT-style, and Encoder-Decoder models—to facilitate the development and evaluation of new methods. The introduction highlights the growing importance of code intelligence in the software development process, emphasizing the need for benchmarks to drive research and innovation. The paper details the datasets, tasks, and baseline systems, and reports experimental results on various tasks, demonstrating the effectiveness of the proposed models. The authors plan to extend the dataset to more programming languages and tasks in the future.CodeXGLUE is a comprehensive benchmark dataset designed to foster machine learning research in program understanding and generation. It includes 14 datasets and 10 tasks across various programming languages, covering clone detection, defect detection, cloze test, code completion, code repair, code-to-code translation, text-code, code-text, and text-text tasks. The dataset supports both supervised and unsupervised learning, making it suitable for a wide range of applications. CodeXGLUE also provides three baseline systems—BERT-style, GPT-style, and Encoder-Decoder models—to facilitate the development and evaluation of new methods. The introduction highlights the growing importance of code intelligence in the software development process, emphasizing the need for benchmarks to drive research and innovation. The paper details the datasets, tasks, and baseline systems, and reports experimental results on various tasks, demonstrating the effectiveness of the proposed models. The authors plan to extend the dataset to more programming languages and tasks in the future.