Given this can produce code when prompted, could it also be used to interpret html from a crawler and then be used to scrape arbitrary URLs and extract structured attributes? Basically like MarkupLM but with massively more token context?