[experiment]
name = "mlm-6"
type = "mlm"


[dataset]
path = "thejosango/nuha-dataset"
dataset_revision = "main"
augment_ratio = 0.0
undersampling_strategy = false


[model]
pretrained_model_name_or_path = "from-scratch"
num_hidden_layers = 8

[training]
num_train_epochs = 25 
warmup_steps = 5e3
lr_scheduler_type = "linear"
learning_rate = 1e-4
per_device_train_batch_size = 128 
per_device_eval_batch_size = 128
gradient_accumulation_steps = 1
weight_decay = 0.01
label_smoothing_factor = 0.1
weighted_loss = false
early_stopping_patience = 5
early_stopping_threshold = 0.005