| # set train hyperparams | |
| unset LD_PRELOAD | |
| export HF_DATASETS_CACHE="/researchdisk/datasets_cache" | |
| export USE_TORCH=0 | |
| python3 run_clm_flax.py \ | |
| --output_dir="./" \ | |
| --model_type="gpt2" \ | |
| --config_name="./" \ | |
| --tokenizer_name="./" \ | |
| --dataset_filepath="/researchdisk/training_dataset_full_deduplicated" \ | |
| --do_train --do_eval \ | |
| --block_size="512" \ | |
| --per_device_train_batch_size="32" \ | |
| --per_device_eval_batch_size="32" \ | |
| --preprocessing_num_workers="1" \ | |
| --distributed_shampoo \ | |
| --learning_rate="1e-4" \ | |
| --warmup_steps="4000" \ | |
| --cosine_decay \ | |
| --overwrite_output_dir \ | |
| --logging_steps="500" \ | |
| --eval_steps="10000" \ | |
| --save_steps="10000" \ | |
| --num_train_epochs="10" \ | |
| --dtype="bfloat16" \ | |
| --push_to_hub \ | |
| --hub_model_id="Finnish-NLP/gpt2-finnish" |