Reward models and dataset for image-text alignment trained with cycle consistency preferences. Project page at https://cyclereward.github.io/