run_trainer.sh

CUDA_VISIBLE_DEVICES=0 python finetune_trainer.py \
 --learning_rate 1e-4 \
 --num_train_epochs 5 \
 --output_dir save/t5-base \
 --model_name_or_path t5-base \
 --weight_decay=0.005 \
 --do_train \
 --do_eval \
 --evaluation_strategy="epoch" \
 --save_strategy="epoch"  \
 --save_total_limit=5 \
 --overwrite_output_dir \
 --gradient_accumulation_steps 16 \
 --per_device_train_batch_size 4 \
 --per_device_eval_batch_size 4 \
 --load_best_model_at_end True \
 --dataset_name src/data_utils/cicero.py \
 --dataset_config_name cicero_nlg


CUDA_VISIBLE_DEVICES=0 python finetune_trainer.py \
 --learning_rate 1e-4 \
 --num_train_epochs 5 \
 --output_dir save/t5-base-cicero-contrast \
 --model_name_or_path t5-base \
 --weight_decay=0.005 \
 --do_train \
 --do_eval \
 --evaluation_strategy="epoch" \
 --save_strategy="epoch"  \
 --save_total_limit=5 \
 --overwrite_output_dir \
 --gradient_accumulation_steps 16 \
 --per_device_train_batch_size 4 \
 --per_device_eval_batch_size 4 \
 --load_best_model_at_end True \
 --dataset_name src/data_utils/cicero.py \
 --dataset_config_name cicero_nlg_contrast \
 --coef_inbatch 0.5 \
 --coef_insample 0.5 \
 --temperature_tau_sample 2.5 \
 --temperature_tau_batch 0.1 \