Extracting memorized pieces of books from open-weight language models