it seems that fitting deltaG's through the web interface where dask fit the pytorch model is (much) worse then running directly without dask