default_train_args.json

{
	"overwrite_output_dir": true,
	"do_train": true,
	"do_eval": true,
	"per_device_train_batch_size": 6,
	"per_device_eval_batch_size": 6,
	"gradient_accumulation_steps": 45,
	"eval_accumulation_steps": 1,
	"evaluation_strategy": "epoch",
	"learning_rate": 5e-05,
	"num_train_epochs": 10,
	"weight_decay": 0.001,
	"max_grad_norm": 1.0,
	"lr_scheduler_type": "linear",
	"warmup_steps": 500,
	"save_strategy": "steps",
	"save_steps": 1000,
	"save_total_limit": 20,
	"save_on_each_node": false,
	"no_cuda": false,
	"seed": 42,
	"fp16": false
}