| #export TOKENIZERS_PARALLELISM=0 | |
| python ./run_mlm_flax_no_accum.py \ | |
| --push_to_hub \ | |
| --output_dir="./" \ | |
| --model_type="big_bird" \ | |
| --config_name="./" \ | |
| --tokenizer_name="./" \ | |
| --max_seq_length="4096" \ | |
| --weight_decay="0.0095" \ | |
| --warmup_steps="10000" \ | |
| --overwrite_output_dir \ | |
| --adam_beta1="0.9" \ | |
| --adam_beta2="0.98" \ | |
| --logging_steps="50" \ | |
| --eval_steps="6000" \ | |
| --num_train_epochs="3"\ | |
| --preprocessing_num_workers="96" \ | |
| --save_steps="15000" \ | |
| --learning_rate="3e-5" \ | |
| --per_device_train_batch_size="1" \ | |
| --per_device_eval_batch_size="1" \ | |
| --save_total_limit="20"\ | |
| --max_eval_samples="4000"\ | |
| --resume_from_checkpoint="./"\ | |
| #--gradient_accumulation_steps="4"\ | |
| #--adafactor \ | |
| #--dtype="bfloat16" \ | |