What AI says about CodeXGLUE

CodeXGLUE is a benchmark dataset and leaderboard for code understanding and generation, comprising 14 datasets across 10 programming-language tasks that span code-code, text-code, code-text, and text-text scenarios. It provides three baseline model families—CodeBERT (BERT-style), CodeGPT (GPT-style), and an encoder-decoder framework—to support solving tasks such as clone detection, defect detection, cloze tests, code completion, translation, and more. To submit results for official evaluation, researchers must generate predictions on the dev set, run official evaluation, generate test predictions, and email the required information (test and dev predictions, team and model details, and citations) to [email protected].