run_sft.sh

torchrun --nproc_per_node 8 scripts/supervised_finetuning.py \
    --model_type baichuan \
    --template_name baichuan-chat \
    --model_name_or_path ./checkpoints/Qilin-Med-Pretrain \
    --train_file_dir ./data/sft \
    --validation_file_dir ./data/sft \
    --deepspeed config/deepspeed_config_zero2.json \
    --per_device_train_batch_size 32 \
    --per_device_eval_batch_size 4 \
    --do_train \
    --do_eval \
    --use_peft True \
    --bf16 \
    --max_train_samples 10000000 \
    --max_eval_samples 1000 \
    --num_train_epochs 1 \
    --learning_rate 2e-5 \
    --warmup_ratio 0.05 \
    --weight_decay 0.05 \
    --logging_strategy steps \
    --logging_steps 10 \
    --eval_steps 50 \
    --evaluation_strategy steps \
    --save_steps 5000 \
    --save_strategy steps \
    --save_total_limit 3 \
    --gradient_accumulation_steps 1 \
    --preprocessing_num_workers 1 \
    --max_source_length 256 \
    --max_target_length 256 \
    --output_dir ./checkpoints/Qilin-Med-SFT \
    --overwrite_output_dir \
    --ddp_timeout 30000 \
    --logging_first_step True \
    --target_modules all \
    --lora_rank 8 \
    --lora_alpha 32 \
    --lora_dropout 0.05 \
    --torch_dtype bfloat16 \
    --device_map auto \
    --ddp_find_unused_parameters False \
    --gradient_checkpointing True