src/run.sh

#!/bin/bash

export LC_ALL=C.UTF-8
export LANG=C.UTF-8

#export MODEL_NAME_OR_PATH=t5-base
export OUTPUT_DIR=/home/saied/code/gpt2-medium-persian
export MODEL_TYPE=gpt2
export CONFIG_NAME=/home/saied/code/gpt2-medium-persian
export TOKENIZER_NAME=/home/saied/code/gpt2-medium-persian

#export TRAIN_FILE=/home/saied/code/data/...csv
#export VALIDATION_FILE=/home/saied/code/data/...csv
#export TEST_FILE=/home/saied/code/data/...csv
export DATASET_NAME=oscar
export DATASET_CONFIG_NAME=unshuffled_deduplicated_fa
export MAX_SEQUENCE_LENGTH=512

#export MAX_TRAIN_SAMPLE=5000
#export MAX_EVAL_SAMPLES=5000

export PER_DEVICE_TRAIN_BATCH_SIZE=16
export PER_DEVICE_EVAL_BATCH_SIZE=16
export NUM_TRAIN_EPOCHS=10.0
export LEARNING_RATE=1e-3
export WARMUP_STEPS=5000
export LOGGING_STEPS=500
export EVAL_STEPS=2500
export SAVE_STEPS=2500

python src/run_clm.py \
    --output_dir="$OUTPUT_DIR"  \
    --model_type="$MODEL_TYPE" \
    --config_name="$CONFIG_NAME" \
    --tokenizer_name="$TOKENIZER_NAME" \
    --dataset_name="$DATASET_NAME" \
    --dataset_config_name="$DATASET_CONFIG_NAME" \
    --max_seq_length="$MAX_SEQUENCE_LENGTH" \
    --per_device_train_batch_size=$PER_DEVICE_TRAIN_BATCH_SIZE \
    --per_device_eval_batch_size=$PER_DEVICE_EVAL_BATCH_SIZE \
    --num_train_epochs=$NUM_TRAIN_EPOCHS \
    --learning_rate=$LEARNING_RATE \
    --warmup_steps=$WARMUP_STEPS \
    --logging_step=$LOGGING_STEPS \
    --eval_steps=$EVAL_STEPS \
    --save_steps=$SAVE_STEPS \
    --do_train \
    --do_eval \
    --overwrite_output_dir \
    --push_to_hub