**Question:** Can the model state be used for the initial condition of the training? **Answer**: Yes, this is actually a good strategy. Instead of running a training submission for one week, you can run it for 1-2 days before retrieving the trained model (checkpoint), analyze it and resume the training by including the trained model in the training container. This approach seems also safer regarding the lose of computational time quota, for example in case of an error that only occurs late in the processing of the training subjects.