python run_mlm_flax_stream.py \ --output_dir="../roberta-debug-32" \ --model_name_or_path="xlm-roberta-base" \ --config_name="./" \ --tokenizer_name="./" \ --dataset_name="NbAiLab/scandinavian" \ --max_seq_length="512" \ --weight_decay="0.01" \ --per_device_train_batch_size="62" \ --per_device_eval_batch_size="62" \ --learning_rate="1e-4" \ --warmup_steps="10000" \ --overwrite_output_dir \ --num_train_steps="1000000" \ --adam_beta1="0.9" \ --adam_beta2="0.98" \ --logging_steps="5000" \ --save_steps="25000" \ --eval_steps="25000" \ --dtype="bfloat16" \ --push_to_hub