The only significant bit of preprocessing by the language model is tokenization. You can see how it works here: https://beta.openai.com/tokenizer