How can a machine learning model have a less restrictive license than the database it was trained on?
I know there was a similar question to this here, but I want to ask more specifically.
I was building a program that a company aims to use commercially later. To help me I used the open source library dlib by Davis King. The models I obtained separately from his github. Generally, the repository they are contained in is licensed with Creative Commons Zero v1.0 Universal. Therefore it should be safe to use for a production environment.
But in contradiction to that there is a note under one contained model, saying this model should not be used ' in a commercial product.', because the dataset excludes commercial use.
I get that, but is it always that the ML model inherits the license of the dataset? Why does Davis King even publish the models under CC0 then? Because also the other models were trained on datasets that might be more restrictive. All in all I am kind of scared to use these pretrained models now, because it is really not clear to me. This extends to the images he scraped from the internet, how am I supposed to know what license they are underlying?
I also have a second example. In the
face-api.js library there is a model contained called SSD Mobilenet v1. It is included in the library that is licensed under MIT. But it was trained on the WIDERFACE database underlying Creative Common License (cc by-nc-nd), so not allowing commercial use. Isn't that totally misleading as there is zero note on this?
I hope my question got clear and there are not too many details.