update readme hiyouga#313

README.md

@@ -149,6 +149,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --save_steps 1000 \
     --learning_rate 5e-5 \
     --num_train_epochs 3.0 \
+    --plot_loss \
     --fp16
 ```
 
@@ -170,6 +171,8 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --do_train \
     --dataset comparison_gpt4_en \
     --finetuning_type lora \
+    --resume_lora_training False \
+    --checkpoint_dir path_to_sft_checkpoint \
     --output_dir path_to_rm_checkpoint \
     --per_device_train_batch_size 4 \
     --gradient_accumulation_steps 4 \
@@ -178,6 +181,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --save_steps 1000 \
     --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
+    --plot_loss \
     --fp16
 ```
 
@@ -201,7 +205,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --save_steps 1000 \
     --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
-    --fp16
+    --plot_loss
 ```
 
 ### Evaluation (BLEU and ROUGE_CHINESE)

README_zh.md

@@ -155,6 +155,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --save_steps 1000 \
     --learning_rate 5e-5 \
     --num_train_epochs 3.0 \
+    --plot_loss \
     --fp16
 ```
 
@@ -176,6 +177,8 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --do_train \
     --dataset comparison_gpt4_zh \
     --finetuning_type lora \
+    --resume_lora_training False \
+    --checkpoint_dir path_to_sft_checkpoint \
     --output_dir path_to_rm_checkpoint \
     --per_device_train_batch_size 4 \
     --gradient_accumulation_steps 4 \
@@ -184,6 +187,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --save_steps 1000 \
     --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
+    --plot_loss \
     --fp16
 ```
 
@@ -196,6 +200,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --do_train \
     --dataset alpaca_gpt4_zh \
     --finetuning_type lora \
+    --resume_lora_training False \
     --checkpoint_dir path_to_sft_checkpoint \
     --reward_model path_to_rm_checkpoint \
     --output_dir path_to_ppo_checkpoint \
@@ -206,7 +211,7 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --save_steps 1000 \
     --learning_rate 1e-5 \
     --num_train_epochs 1.0 \
-    --fp16
+    --plot_loss
 ```
 
 ### 指标评估（BLEU分数和汉语ROUGE分数）

examples/covid_doctor.md

@@ -70,13 +70,15 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --dataset comparison_gpt4_zh \
     --dataset_dir data \
     --finetuning_type lora \
+    --resume_lora_training False \
+    --checkpoint_dir covid/sft \
     --output_dir covid/rm \
     --overwrite_cache \
-    --per_device_train_batch_size 4 \
-    --per_device_eval_batch_size 4 \
+    --per_device_train_batch_size 2 \
+    --per_device_eval_batch_size 2 \
     --gradient_accumulation_steps 8 \
-    --max_source_length 512 \
-    --max_target_length 512 \
+    --max_source_length 1024 \
+    --max_target_length 128 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --save_steps 100 \
@@ -109,20 +111,20 @@ CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
     --dataset covid_train,covid_dev \
     --dataset_dir data/covid \
     --finetuning_type lora \
+    --resume_lora_training False \
     --checkpoint_dir covid/sft \
     --reward_model covid/rm \
     --output_dir covid/ppo \
     --overwrite_cache \
     --per_device_train_batch_size 2 \
-    --gradient_accumulation_steps 4 \
+    --gradient_accumulation_steps 8 \
     --max_source_length 256 \
     --max_target_length 128 \
     --lr_scheduler_type cosine \
     --logging_steps 10 \
     --save_steps 100 \
     --learning_rate 1e-5 \
     --num_train_epochs 5.0 \
-    --resume_lora_training False \
     --plot_loss
 ```
 

examples/quantized_finetune_with_local_model.sh

@@ -16,6 +16,6 @@ CUDA_VISIBLE_DEVICES=0 python ../src/train_bash.py \
     --save_steps 1000 \
     --learning_rate 5e-5 \
     --num_train_epochs 3.0 \
-    --quantization_bit 8 \
+    --quantization_bit 4 \
     --plot_loss \
     --fp16

examples/train_rm.sh

@@ -6,6 +6,8 @@ CUDA_VISIBLE_DEVICES=0 python ../src/train_bash.py \
     --dataset comparison_gpt4_zh \
     --dataset_dir ../data \
     --finetuning_type lora \
+    --resume_lora_training False \
+    --checkpoint_dir path_to_sft_checkpoint \
     --output_dir path_to_rm_checkpoint \
     --overwrite_cache \
     --per_device_train_batch_size 2 \