Datasets for Reinforced Adversarial Learning: training reasoning models from reference answers alone.