Stratified train-test split

The problem, from first principles

Suppose 5 percent of your examples are positive. A random split can, by chance, put almost all of them in one side, and then your test set says little about how the model handles the rare class. A stratified split fixes the class proportions on both sides: each class is split separately, so the test set keeps the same mix as the full dataset.

Implement stratified_train_test_split(y, test_fraction=0.25, seed=0), which returns (train_indices, test_indices) as sorted integer arrays. Each class is shuffled with a seeded generator and round(test_fraction * class_count) of its members go to the test set.

Constraints

• Every index appears in exactly one of the two outputs.
• For each class, the number of test examples is round(test_fraction * class_count).
• test_fraction must lie in [0, 1], otherwise raise ValueError.
• The same seed gives the same split.
• Both outputs are sorted.

Hints

Open one at a time. Each gives away a little more than the last.