Fine-tuning language models to find agreement among humans with diverse preferences