And what about the case when an internal and proprietary company tool was used to train the model (e.g., an internal wrapper on top of TensorFlow)?
May I ask why the model accompanied with the inference code - which fully allows you to reproduce our results on the evaluation set - are not...