The value responses from 1,000 test subjects were used to tune a more democratic large language model.
