Try max_depth = 3, 5, 10 on the training data and pick the one with the highest training accuracy

Hold out a final test set, then use k-fold cross-validation on the remaining data to evaluate max_depth values from 1 to 30; select the depth with the best average CV score, then report performance on the held-out test set

Use the full dataset with a single 50/50 split for each depth value

Use cross-validation to test max_depth values and report the best CV score as the final performance estimate