Questions about building predictive models

[GA-1101]

When we choose a variety of cell types (e.g. batch_key is a cell type), whether SCGEN constructs a sub-model according to batch_key, or build a unified model based on all Train data? In the subsequent Prediction process, what is the content corresponding to the restrict_arithmetic_to parameter? Does it mean that the constructed sub-models are selected separately according to the parameter content, how should we understand this parameter?