Use package https://github.com/toon-format/toon for both golden dataset and output files. This can reduce token usage by 60%.