Training data is the set of samples, labelled or not, used to train the algorithm; its size and the choice of features govern how well the trained model, which the standard defines as the result of establishing or improving model parameters from training data, fits the target. To assess the model the acquired data is split into training, validation and test sets: validation data tunes hyperparameters during and after training, test data confirms the model learned what it should, and neither is shown to the model during training, nor is test data used during tuning, otherwise the model can remember answers and performance is overstated. Cross-validation reuses samples for training and validation to emulate a larger dataset; where data is scarce, a two-way split is made and validation sets are generated from the training portion by bootstrapping or cross-validation.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.