This paper introduces the BMA-Measure, a new metric designed to evaluate the quality of the training set. While most previous work has focused on evaluating classification algorithms, the importance of the representativeness of the example base for the accuracy of the classification process and the reliability of the generated models has often been overlooked. The BMA-Measure aims to fill this gap by strengthening existing measures and helping experts better interpret results, optimize evaluation reliability and robustness, and precisely identify gaps. Developed from theoretical criteria for generating an optimal training set, the BMA-Measure proposes a simple and appropriate approach for each sub-problem before combining all these metrics into one. This metric is particularly important in sensitive sectors such as digital health, where the reliability of learning models is crucial to ensuring accurate and unbiased predictions. By assessing the representativeness of the example base, the BMA-Measure contributes to reinforcing this confidence by demonstrating the quality and impartiality of learning models.