roberta-swahili / run.sh
Patrick von Platen
Saving weights and logs of step 1000
2b6c375
raw
history blame
630 Bytes
./run_mlm_flax.py \
--output_dir="./" \
--model_type="roberta" \
--config_name="./" \
--tokenizer_name="./" \
--dataset_name="mc4" \
--dataset_config_name="sw" \
--max_seq_length="128" \
--weight_decay="0.025" \
--per_device_train_batch_size="128" \
--per_device_eval_batch_size="128" \
--learning_rate="3e-4" \
--warmup_steps="500" \
--overwrite_output_dir \
--num_train_epochs="15" \
--adam_beta1="0.9" \
--adam_beta2="0.98" \
--logging_steps="250" \
--save_steps="1000" \
--eval_steps="1000" \
--preprocessing_num_workers="96" \
--push_to_hub