GPT model with UL2 objective (Did not perform better than causal language modeling objective)