As discussed in #14300 (comment) it might be useful to add the d2_deviance_score that is a generalization of r2_score to non normal error distributions. In the same way as mean_tweedie_deviance (introduced #14263) is a generalization of mean_squared_error .