Learning the rules of peptide self-assembly through data mining with large language models