A curious question. Would it be possible to build LLMs off of bytecode / compiled code, and then translate it back to different programming language codes?

Is that a viable way to get more fine-tuned outputs from these models?