are there any suggestions for this? I am not sure if hidden[:, 0, :] makes sense (since no [CLS] token in T5) but I found that using hidden[:,0,:] is yielding better results than torch.mean(hidden_states, dim=1). Any suggestions on whats the best way to do this in T5Encoder?