Tabnine trains on open-source code with permissive licenses. It is a combination of our own GPT models, which follow GPT- 3.5 architecture but are organized as a collection of several models with sizes up to 15B parameters.
See the entire training set repo list here.