change: new train yaml

batch_run.sh

@@ -1,9 +1,9 @@
-FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_1.yaml | tee results/lora_sft/Llama2-7B/llama2_lora_sft_1.txt
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Llama2-7B/llama2_lora_sft_1.yaml | tee results/lora_sft_2/Llama2-7B/llama2_lora_sft_1.txt
-FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_2.yaml | tee results/lora_sft/Llama2-7B/llama2_lora_sft_2.txt
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Llama2-7B/llama2_lora_sft_2.yaml | tee results/lora_sft_2/Llama2-7B/llama2_lora_sft_2.txt
-FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_3.yaml | tee results/lora_sft/Llama2-7B/llama2_lora_sft_3.txt
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Llama2-7B/llama2_lora_sft_3.yaml | tee results/lora_sft_2/Llama2-7B/llama2_lora_sft_3.txt
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_1_single.yaml | tee results/lora_sft/Llama2-7B/llama2_lora_sft_1_single.txt
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Llama2-7B/llama2_lora_sft_1_single.yaml | tee results/lora_sft_2/Llama2-7B/llama2_lora_sft_1_single.txt
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_2_single.yaml | tee results/lora_sft/Llama2-7B/llama2_lora_sft_2_single.txt
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Llama2-7B/llama2_lora_sft_2_single.yaml | tee results/lora_sft_2/Llama2-7B/llama2_lora_sft_2_single.txt
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_3_single.yaml | tee results/lora_sft/Llama2-7B/llama2_lora_sft_3_single.txt
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Llama2-7B/llama2_lora_sft_3_single.yaml | tee results/lora_sft_2/Llama2-7B/llama2_lora_sft_3_single.txt
 
 # FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_1.yaml | tee results/inference/Llama2-7B/llama2_predict_1.txt
 # FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_2.yaml | tee results/inference/Llama2-7B/llama2_predict_2.txt
@@ -12,26 +12,26 @@ CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_
 # CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_2_single.yaml | tee results/inference/Llama2-7B/llama2_predict_2_single.txt
 # CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_3_single.yaml | tee results/inference/Llama2-7B/llama2_predict_3_single.txt
 
-# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1.yaml | tee results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1.yaml | tee results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1.txt
-# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2.yaml | tee results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_2.yaml | tee results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_2.txt
-# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3.yaml | tee results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_3.yaml | tee results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_3.txt
-# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1_single.yaml | tee results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1_single.yaml | tee results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1_single.txt
-# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2_single.yaml | tee results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_2_single.yaml | tee results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_2_single.txt
-# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3_single.yaml | tee results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_3_single.yaml | tee results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_3_single.txt
 
-# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1.yaml | tee results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_1.yaml | tee results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_1.txt
-FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2.yaml | tee results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_2.yaml | tee results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_2.txt
-FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3.yaml | tee results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_3.yaml | tee results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_3.txt
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1_single.yaml | tee results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_1_single.yaml | tee results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_1_single.txt
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2_single.yaml | tee results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_2_single.yaml | tee results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_2_single.txt
-CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3_single.yaml | tee results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_3_single.yaml | tee results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_3_single.txt
 
-# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_1.yaml | tee results/lora_sft/Qwen-7B/Qwen_lora_sft_1.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Qwen-7B/Qwen_lora_sft_1.yaml | tee results/lora_sft_2/Qwen-7B/Qwen_lora_sft_1.txt
-# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_2.yaml | tee results/lora_sft/Qwen-7B/Qwen_lora_sft_2.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Qwen-7B/Qwen_lora_sft_2.yaml | tee results/lora_sft_2/Qwen-7B/Qwen_lora_sft_2.txt
-# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_3.yaml | tee results/lora_sft/Qwen-7B/Qwen_lora_sft_3.txt
+# FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft_2/Qwen-7B/Qwen_lora_sft_3.yaml | tee results/lora_sft_2/Qwen-7B/Qwen_lora_sft_3.txt
-# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_1_single.yaml | tee results/lora_sft/Qwen-7B/Qwen_lora_sft_1_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Qwen-7B/Qwen_lora_sft_1_single.yaml | tee results/lora_sft_2/Qwen-7B/Qwen_lora_sft_1_single.txt
-# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_2_single.yaml | tee results/lora_sft/Qwen-7B/Qwen_lora_sft_2_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Qwen-7B/Qwen_lora_sft_2_single.yaml | tee results/lora_sft_2/Qwen-7B/Qwen_lora_sft_2_single.txt
-# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_3_single.yaml | tee results/lora_sft/Qwen-7B/Qwen_lora_sft_3_single.txt
+# CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft_2/Qwen-7B/Qwen_lora_sft_3_single.yaml | tee results/lora_sft_2/Qwen-7B/Qwen_lora_sft_3_single.txt
 
 # FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_1.yaml | tee results/inference/Qwen-7B/Qwen_predict_1.txt
 # FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_2.yaml | tee results/inference/Qwen-7B/Qwen_predict_2.txt

results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: baichuan
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: baichuan
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_2.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: baichuan
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_2_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: baichuan
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_3.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: baichuan
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_3_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: baichuan
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_1.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/chatglm/data
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: chatglm2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_1_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/chatglm/data
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: chatglm2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_2.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/chatglm/data
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: chatglm2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_2_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/chatglm/data
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: chatglm2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_3.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/chatglm/data
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: chatglm2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/ChatGLM2-6B/ChatGLM2_lora_sft_3_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/chatglm/data
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: chatglm2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Llama2-7B/llama2_lora_sft_1.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: llama2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Llama2-7B/llama2_lora_sft_1
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+fp16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Llama2-7B/llama2_lora_sft_1_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: llama2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Llama2-7B/llama2_lora_sft_1_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+fp16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Llama2-7B/llama2_lora_sft_2.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: llama2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Llama2-7B/llama2_lora_sft_2
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+fp16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Llama2-7B/llama2_lora_sft_2_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: llama2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Llama2-7B/llama2_lora_sft_2_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+fp16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Llama2-7B/llama2_lora_sft_3.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: llama2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Llama2-7B/llama2_lora_sft_3
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+fp16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Llama2-7B/llama2_lora_sft_3_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: llama2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Llama2-7B/llama2_lora_sft_3_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+fp16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Qwen-7B/Qwen_lora_sft_1.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/qwen
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: qwen
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Qwen-7B/Qwen_lora_sft_1
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Qwen-7B/Qwen_lora_sft_1_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/qwen
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: qwen
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Qwen-7B/Qwen_lora_sft_1_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Qwen-7B/Qwen_lora_sft_2.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/qwen
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: qwen
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Qwen-7B/Qwen_lora_sft_2
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Qwen-7B/Qwen_lora_sft_2_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/qwen
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: qwen
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Qwen-7B/Qwen_lora_sft_2_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Qwen-7B/Qwen_lora_sft_3.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/qwen
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: qwen
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Qwen-7B/Qwen_lora_sft_3
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/lora_sft_2/Qwen-7B/Qwen_lora_sft_3_single.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: ../../llm/qwen
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: qwen
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft/Qwen-7B/Qwen_lora_sft_3_single
+logging_steps: 3
+save_steps: 100
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+max_steps: 1000
+include_num_input_tokens_seen: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500