Best DAG doesn't seem to be saved on it's own

I notice that in Trainer.train() self.save_model() is called at certain times to save all of the shared weights in the "super graph" (my terminology), but I don't see that the best dag is tracked such that at the end of train() we have the best DAG/model RNN for the PTB task. save_model() saves all of the weights for the entire shared weight space, but doesn't show which DAG (which sub-graph of the larger graph) represents the best RNN constructed during training - unless I'm missing something?

Start by locating Trainer.train() and save_model(), then trace how the PTB task identifies the best DAG or model RNN during training. Determine what saved output should identify the best sub-graph in addition to the shared weights, and verify that the resulting artifact can distinguish that DAG at the end of train().

Written by the indexing model from the issue text.