Stratified Split

from sklearn.model_selection import train_test_split
X_tr, X_te, y_tr, y_te = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42
)

Why it matters: With class imbalance, a naive split can put all minority examples in train or test. Stratify preserves class ratios across splits — otherwise your metrics are noise.