commands.txt

# note CUDA_VISIBLE_DEVICES sets which of the GPUs it is allowed to use. Given the GPUs, it automatically selects them.

# supervised pretraining 
CUDA_VISIBLE_DEVICES=1,2,3 python pre_training_script.py --subset None --dataset_name "csv" --file_path "input_data/med_data.csv" --max_steps 200 --save_steps 20 --logging_steps 1 --log_freq 1 --size_valid_set 200 --shuffle_buffer 200 --seed 1

# DPO training
CUDA_VISIBLE_DEVICES=1,2,3 python training_script.py --model_name_or_path sft/final_checkpoint/ --dataset_name "csv" --file_path "input_data/med_data.csv" --per_device_train_batch_size 1 --gradient_accumulation_steps 16 --max_steps 200 --save_steps 20 --logging_steps 1 --output_dir ./dpo_1

# testing
CUDA_VISIBLE_DEVICES=5,6,7 python testing_script.py --model_name_or_path "dpo/final_checkpoint/"

# writing all checkpoint outputs to csv
CUDA_VISIBLE_DEVICES=1,2,3 python checkpoints_to_csv.py --question_file input_data/med_questions.csv