Jumps in training performance due to file changes
It can be observed that the train performance for the ORCA4 MC (first batch, v5.4) jumps when the file is changed. It is unclear to me where this could come from. Maybe it is worth investigating further as it may indicate a feature of the MC production? I dont know^^. When randomizing the order if the runs considered for a file, this behavior is gone. In the contrib files from Michael I thus suggest the change in the following merge request.