| ./run_mlm_flax_stream.py \ | |
| --output_dir="./" \ | |
| --model_type="roberta" \ | |
| --config_name="./" \ | |
| --tokenizer_name="./" \ | |
| --model_name_or_path="./" | |
| --dataset_name="pere/norwegian_colossal_corpus_v2_short100k" \ | |
| --max_seq_length="128" \ | |
| --weight_decay="0.01" \ | |
| --per_device_train_batch_size="128" \ | |
| --per_device_eval_batch_size="128" \ | |
| --learning_rate="2.0253e-4" \ | |
| --warmup_steps="0" \ | |
| --overwrite_output_dir \ | |
| --cache_dir /mnt/disks/flaxdisk/cache/ \ | |
| --num_train_steps="1650000" \ | |
| --adam_beta1="0.9" \ | |
| --adam_beta2="0.98" \ | |
| --logging_steps="10000" \ | |
| --save_steps="100000" \ | |
| --eval_steps="50000" \ | |
| --preprocessing_num_workers 96 \ | |
| --adafactor \ | |
| --push_to_hub | |