run.sh

export HF_PROJECT="t5-v1_1-base-dutch-english-cased"
export DATASET="yhavinga/mc4_nl_cleaned" # Name of the dataset in the Huggingface Hub
export DATASET_CONFIG="small_en_nl" # Config of the dataset in the Huggingface Hub
export DATASET_SPLIT="train" # Split to use for training tokenizer and model
export CONFIG_NAME="yhavinga/t5-v1.1-base-dutch-cased"
export TOKENIZER_NAME="yhavinga/net5-v1.1-base-cased-500"
export MODEL_PATH="${HOME}/data/${HF_PROJECT}" # Path to the model


python3 ../train/run_t5_mlm_flax_pmap.py \
    --output_dir="${MODEL_PATH}" \
    --model_type="t5" \
    --config_name="${CONFIG_NAME}" \
    --tokenizer_name="${TOKENIZER_NAME}" \
    --auth_token="$(cat ~/.huggingface/token)" \
    --preprocessing_num_workers="96" \
    --do_train --do_eval \
    --dataset_name="${DATASET}" \
    --dataset_config_name="${DATASET_CONFIG}" \
    --max_seq_length="512" \
    --per_device_train_batch_size="16" \
    --per_device_eval_batch_size="16" \
    --optim="adafactor" \
    --learning_rate="0.005" \
    --lr_decay="exponential" \
    --lr_transition_steps="300000" \
    --lr_decay_rate="0.7" \
    --lr_staircase="false" \
    --overwrite_output_dir \
    --num_train_epochs="10" \
    --logging_steps="200" \
    --save_steps="10000" \
    --eval_steps="1250" \
    --warmup_steps="10000" \
    --validation_split_count="15000" \
    --wandb_project="t5-v1_1-dutch-english" \
    --wandb_job_type="pmap" \
    --resume_from_checkpoint="${MODEL_PATH}"