I've recently come to learn that filling the context window up completely makes the results worse and the llm struggles to acquire the tokens in the middle of the prompt. So my question is : Are there any studies on this and what is the appropriate fill for a context window ?