I suppose giving out model to try out and get the training data in return is a good way for the inference providers to learn from data.