Yes and no. Suppose for example you give MNIST to a SVM and fit the model, then test it only on 0 and 1 digits, which are generally well discriminized, you'll g

Yes and no. Suppose for example you give MNIST to a SVM and fit the model, then test it only on 0 and 1 digits, which are generally well discriminized, you'll get almost 100% accuracy in test, whereas 97% or less in training. (ye probably need some preprocessing, like using PacMAP or UMAP or whatever but the point is the same) However, that's just because I decided the right data to test it onto. So, you can't really say much on a model using that definition.