Well if the LLM is closed and proprietary, there is no insight into how training data is even used. It's just a black-box we have to use blindly and 'hope' the designers are using a blend of fine-tuning coupled with better training data.