This isn’t even a question of training data, thy fed the full git source code directly to the llm.
[1]: https://malus.sh/