We present CodeBERT, a bimodal pre-trained model for programming language(PL) and nat-ural language (NL). CodeBERT learns general-purposerepresentations that support downstream NL-PL applications such as naturallanguage codesearch, code documentation generation, etc. We develop CodeBERTwith Transformer-based neural architecture, and train it with a hybridobjective function that incorporates the pre-training task of replaced tokendetection, which is to detect plausible alternatives sampled from generators.This enables us to utilize both bimodal data of NL-PL pairs and unimodal data,where the former provides input tokens for model training while the latterhelps to learn better generators. We evaluate CodeBERT on two NL-PLapplications by fine-tuning model parameters. Results show that CodeBERTachieves state-of-the-art performance on both natural language code search andcode documentation generation tasks. Furthermore, to investigate what type ofknowledge is learned in CodeBERT, we construct a dataset for NL-PL probing, andevaluate in a zero-shot setting where parameters of pre-trained models arefixed. Results show that CodeBERT performs better than previous pre-trainedmodels on NL-PL probing.