Dataset from the MIST framework: Mutual Information estimation via Supervised Training

This dataset consists of meta-datasets built for supervised training and evaluation of meta-learned mutual information (MI) estimators. Using the BMI library, we generate complex distributions with known MI by applying invertible transformations to simple base distributions. The base distribution families are split into disjoint training and testing sets, ensuring different supports between the training and test meta-distributions. Each meta-datapoint contains paired samples drawn from a joint distribution and their corresponding MI value. The datasets cover dimensions from 2 to 32 and sample sizes between 10 and 500. The training meta-dataset M_train includes 16 base distribution families and 625k meta-datapoints. Two test sets are provided: a small one (M_test, 2.3k points) for slow baselines, and an extended one (M_test_extended, 806k points) that includes both seen and unseen distribution families.