Text this: Learning atomic forces from uncertainty-calibrated adversarial attacks