include 't5x/examples/t5/mt5/large.gin' include 'pretrain_cont.gin' #include 't5x/configs/runs/pretrain.gin' #iinclude 't5x/configs/runs/finetune.gin' # Register necessary SeqIO Tasks/Mixtures. import t5.data.mixtures import tasks MIXTURE_OR_TASK_NAME = "balanced_bokmaal_nynorsk_span_corruption_stream" TASK_FEATURE_LENGTHS = {"inputs": 512, "targets": 512} TRAIN_STEPS = 1_700_000 DROPOUT_RATE = 0.0 # Changed from the default since T5-1.1 recomments this. INITIAL_CHECKPOINT_PATH = "gs://nb-t5x-us-central2/norwegian_NCC_plus_English_t5x_large/checkpoint_1500000" PjitPartitioner.num_partitions = 2 utils.SaveCheckpointConfig.period = 5000 utils.SaveCheckpointConfig.keep = 3