LLMs trained on code; interpretability