python run_seq2seq_flax.py \ --dataset_repo_or_path dalle-mini/encoded \ --train_file **/train/CC3M/*.jsonl \ --validation_file **/valid/*/*.jsonl \ --len_train 129847128 \ --len_eval 157312 \ --streaming \ --output_dir output \ --per_device_train_batch_size 16 \ --per_device_eval_batch_size 16 \ --preprocessing_num_workers 80 \ --warmup_steps 125 \ --gradient_accumulation_steps 8 \ --do_train \ --do_eval \ --adafactor \ --num_train_epochs 1 \ --max_train_samples 10000 \ --learning_rate 0.005