At the moment the way prediction processors (aka machine learning backends) can return data to moodle is quite limited, they return the accuracy, standard deviation of all evaluations and an 'info' array that is dumped into the log information screen.

I think that using a generic 'info' array for general stuff is ok, but we should improve this. e.g. The tensorflow-based model returns a CLI command to view something like https://www.tensorflow.org/get_started/summaries_and_tensorboard.