Aggregate data run /w permutation data

[dennlinger]

Also worth thinking about aggregating the available German data sets, and using permutations of source sentences during training to weaken the positional bias in models to see how they influence training generalization.