Scrubbing tokens from source code is not enough, as shown by the publishing of a Python Software Foundation access token with administrator privileges to a container image on Docker Hub.

This variant has been trained on an additional 100 billion tokens of Python code, making it more fluent and accurate in generating Python code.