start_train.sh

# set train hyperparams
unset LD_PRELOAD
export HF_DATASETS_CACHE="/researchdisk/datasets_cache"
export USE_TORCH=0
python3 run_mlm_flax.py \
    --output_dir="./" \
    --model_name_or_path="./" \
    --config_name="./" \
    --tokenizer_name="./" \
    --dataset_filepath="/researchdisk/training_dataset_full_deduplicated" \
    --max_seq_length="512" \
    --pad_to_max_length \
    --preprocessing_num_workers="64" \
    --per_device_train_batch_size="8" \
    --per_device_eval_batch_size="8" \
    --adam_beta1="0.9" \
    --adam_beta2="0.98" \
    --adam_epsilon="1e-6" \
    --learning_rate="2e-4" \
    --weight_decay="0.01" \
    --warmup_steps="2500" \
    --overwrite_output_dir \
    --num_train_epochs="4" \
    --save_steps="10000" \
    --eval_steps="10000" \
    --logging_steps="500" \
    --dtype="bfloat16" \
    --adafactor \
    --push_to_hub \
    --hub_model_id="Finnish-NLP/roberta-large-wechsel-finnish"