Hyperparameters batch_size = 96 n_epochs = 2 base_LM_model = "roberta-base" max_seq_len = 386 learning_rate = 3e-5 lr_schedule = LinearWarmup warmup_proportion = 0.2 doc_stride=128 max_query_length=64