This is really a fantastic analysis of a dataset, and it's something that should be a mandatory form of smokescreen before proceeding with actual model training, in every organization or research group. Whether you are using public datasets, buying 3rd party data, doing in-house data collection and annotation, or paying someone to do it for you, you must check for class imbalance and over/under representation within your data - inevitably human biases will creep in. Ultimately you have to evaluate whether this data distribution is compatible with your target data distribution your model will be applied on in production. Doing this post hoc is a real pain.