Work out train test split size instantly with clear inputs, formula shown and shareable results.
A three-way split gives the model data to learn from, a validation set for tuning decisions, and a test set touched only once at the end. The validation share is whatever the training and test shares leave over. Row counts matter more than percentages: a 15 percent test set of 300 rows is 45 examples, far too few to separate two models with confidence.
Split sizing
train = rows x train share; test = rows x test share; validation = rows - train - test
Every tuning decision made against a set leaks information into it. Holding a genuinely untouched test set is the only way to get an unbiased estimate of generalisation.
Prefer k-fold cross-validation for model selection and keep one modest held-out test set. Repeated fitting uses every row for training at some point.