start_train.sh

unset LD_PRELOAD
python3 run_mlm_flax.py \
    --output_dir="./" \
    --model_type="roberta" \
    --config_name="./" \
    --tokenizer_name="./" \
    --dataset_filepath="/researchdisk1/data/training_data_full" \
    --save_tokenized_dataset_filepath="/researchdisk1/data/training_data_full_tokenized_128" \
    --max_seq_length="128" \
    --pad_to_max_length \
    --preprocessing_num_workers="96" \
    --per_device_train_batch_size="128" \
    --per_device_eval_batch_size="128" \
    --learning_rate="3e-4" \
    --warmup_steps="1000" \
    --overwrite_output_dir \
    --num_train_epochs="3" \
    --save_strategy="steps" \
    --save_steps="10000" \
    --save_total_limit="5" \
    --eval_steps="10000" \
    --logging_steps="1000" \
    --dtype="bfloat16" \
    --push_to_hub