The tl;dr seems to be: Tell a LLM to create pairs of questions and answers based off of a document, and fine-tune on that data. Does the model answer questions from the article that weren't generated in advance?