However, we recommend using the sep_chains=True argument in the wrapper class for maximum performance on downstream tasks. This gets the sequence-level embedding for all sequences, and returns it ...