|
|
|
@ -41,7 +41,7 @@ python finetune.py \
|
|
|
|
--data_path $DATA \
|
|
|
|
--data_path $DATA \
|
|
|
|
--bf16 True \
|
|
|
|
--bf16 True \
|
|
|
|
--output_dir output_qwen \
|
|
|
|
--output_dir output_qwen \
|
|
|
|
--num_train_epochs 5 \
|
|
|
|
--num_train_epochs 16 \
|
|
|
|
--per_device_train_batch_size 2 \
|
|
|
|
--per_device_train_batch_size 2 \
|
|
|
|
--per_device_eval_batch_size 1 \
|
|
|
|
--per_device_eval_batch_size 1 \
|
|
|
|
--gradient_accumulation_steps 8 \
|
|
|
|
--gradient_accumulation_steps 8 \
|
|
|
|
@ -56,7 +56,7 @@ python finetune.py \
|
|
|
|
--lr_scheduler_type "cosine" \
|
|
|
|
--lr_scheduler_type "cosine" \
|
|
|
|
--logging_steps 1 \
|
|
|
|
--logging_steps 1 \
|
|
|
|
--report_to "none" \
|
|
|
|
--report_to "none" \
|
|
|
|
--model_max_length 512 \
|
|
|
|
--model_max_length 3072 \
|
|
|
|
--lazy_preprocess True \
|
|
|
|
--lazy_preprocess True \
|
|
|
|
--gradient_checkpointing \
|
|
|
|
--gradient_checkpointing \
|
|
|
|
--use_lora
|
|
|
|
--use_lora
|
|
|
|
|
