add: add batch run scripts

batch_run.sh

@@ -0,0 +1,56 @@
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_1.yaml > train results/lora_sft/Llama2-7B/llama2_lora_sft_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_2.yaml > train results/lora_sft/Llama2-7B/llama2_lora_sft_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_3.yaml > train results/lora_sft/Llama2-7B/llama2_lora_sft_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_1_single.yaml > train results/lora_sft/Llama2-7B/llama2_lora_sft_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_2_single.yaml > train results/lora_sft/Llama2-7B/llama2_lora_sft_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Llama2-7B/llama2_lora_sft_3_single.yaml > train results/lora_sft/Llama2-7B/llama2_lora_sft_3_single.log
+
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_1.yaml > train results/inference/Llama2-7B/llama2_predict_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_2.yaml > train results/inference/Llama2-7B/llama2_predict_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_3.yaml > train results/inference/Llama2-7B/llama2_predict_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_1_single.yaml > train results/inference/Llama2-7B/llama2_predict_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_2_single.yaml > train results/inference/Llama2-7B/llama2_predict_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Llama2-7B/llama2_predict_3_single.yaml > train results/inference/Llama2-7B/llama2_predict_3_single.log
+
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1.yaml > train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2.yaml > train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3.yaml > train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1_single.yaml > train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2_single.yaml > train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3_single.yaml > train results/lora_sft/Baichuan2-7B/Baichuan2_lora_sft_3_single.log
+
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1.yaml > train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2.yaml > train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3.yaml > train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1_single.yaml > train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2_single.yaml > train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3_single.yaml > train results/lora_sft/ChatGLM2-6B/ChatGLM2_lora_sft_3_single.log
+
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_1.yaml > train results/lora_sft/Qwen-7B/Qwen_lora_sft_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_2.yaml > train results/lora_sft/Qwen-7B/Qwen_lora_sft_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_3.yaml > train results/lora_sft/Qwen-7B/Qwen_lora_sft_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_1_single.yaml > train results/lora_sft/Qwen-7B/Qwen_lora_sft_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_2_single.yaml > train results/lora_sft/Qwen-7B/Qwen_lora_sft_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/lora_sft/Qwen-7B/Qwen_lora_sft_3_single.yaml > train results/lora_sft/Qwen-7B/Qwen_lora_sft_3_single.log
+
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_1.yaml > train results/inference/Qwen-7B/Qwen_predict_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_2.yaml > train results/inference/Qwen-7B/Qwen_predict_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_3.yaml > train results/inference/Qwen-7B/Qwen_predict_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_1_single.yaml > train results/inference/Qwen-7B/Qwen_predict_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_2_single.yaml > train results/inference/Qwen-7B/Qwen_predict_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Qwen-7B/Qwen_predict_3_single.yaml > train results/inference/Qwen-7B/Qwen_predict_3_single.log
+
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/ChatGLM2-6B/ChatGLM2_predict_1.yaml > train results/inference/ChatGLM2-6B/ChatGLM2_predict_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/ChatGLM2-6B/ChatGLM2_predict_2.yaml > train results/inference/ChatGLM2-6B/ChatGLM2_predict_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/ChatGLM2-6B/ChatGLM2_predict_3.yaml > train results/inference/ChatGLM2-6B/ChatGLM2_predict_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/ChatGLM2-6B/ChatGLM2_predict_1_single.yaml > train results/inference/ChatGLM2-6B/ChatGLM2_predict_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/ChatGLM2-6B/ChatGLM2_predict_2_single.yaml > train results/inference/ChatGLM2-6B/ChatGLM2_predict_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/ChatGLM2-6B/ChatGLM2_predict_3_single.yaml > train results/inference/ChatGLM2-6B/ChatGLM2_predict_3_single.log
+
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Baichuan2-7B/Baichuan2_predict_1.yaml > train results/inference/Baichuan2-7B/Baichuan2_predict_1.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Baichuan2-7B/Baichuan2_predict_2.yaml > train results/inference/Baichuan2-7B/Baichuan2_predict_2.log
+FORCE_TORCHRUN=1 llamafactory-cli train results/inference/Baichuan2-7B/Baichuan2_predict_3.yaml > train results/inference/Baichuan2-7B/Baichuan2_predict_3.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Baichuan2-7B/Baichuan2_predict_1_single.yaml > train results/inference/Baichuan2-7B/Baichuan2_predict_1_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Baichuan2-7B/Baichuan2_predict_2_single.yaml > train results/inference/Baichuan2-7B/Baichuan2_predict_2_single.log
+CUDA_VISIBLE_DEVICES=0 llamafactory-cli train results/inference/Baichuan2-7B/Baichuan2_predict_3_single.yaml > train results/inference/Baichuan2-7B/Baichuan2_predict_3_single.log
+

results/full/Llama2-7B/llama2_full_1.yaml

@@ -0,0 +1,39 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z3_config.json
+
+### dataset
+dataset: belle_1m
+template: llama2
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/full/Llama2-7B/llama2_full_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-5
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/inference/Baichuan2-7B/Baichuan2_predict_1.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: baichuan
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Baichuan2-7B/Baichuan2_predict_1
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Baichuan2-7B/Baichuan2_predict_1_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: baichuan
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Baichuan2-7B/Baichuan2_predict_1_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Baichuan2-7B/Baichuan2_predict_2.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: baichuan
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Baichuan2-7B/Baichuan2_predict_2
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Baichuan2-7B/Baichuan2_predict_2_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: baichuan
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Baichuan2-7B/Baichuan2_predict_2_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Baichuan2-7B/Baichuan2_predict_3.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: baichuan
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Baichuan2-7B/Baichuan2_predict_3
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Baichuan2-7B/Baichuan2_predict_3_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: baichuan
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Baichuan2-7B/Baichuan2_predict_3_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/ChatGLM2-6B/ChatGLM2_predict_1.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: chatglm2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/ChatGLM2-6B/ChatGLM2_predict_1
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/ChatGLM2-6B/ChatGLM2_predict_1_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: chatglm2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/ChatGLM2-6B/ChatGLM2_predict_1_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/ChatGLM2-6B/ChatGLM2_predict_2.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: chatglm2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/ChatGLM2-6B/ChatGLM2_predict_2
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/ChatGLM2-6B/ChatGLM2_predict_2_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: chatglm2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/ChatGLM2-6B/ChatGLM2_predict_2_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/ChatGLM2-6B/ChatGLM2_predict_3.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: chatglm2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/ChatGLM2-6B/ChatGLM2_predict_3
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/ChatGLM2-6B/ChatGLM2_predict_3_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: chatglm2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/ChatGLM2-6B/ChatGLM2_predict_3_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Llama2-7B/llama2_predict_1.yaml

@@ -13,7 +13,7 @@ overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: ./results/inference/Llama2-7B/Llama2-7B_inference_08_15_11_06
+output_dir: ./results/inference/Llama2-7B/llama2_predict_1
 overwrite_output_dir: true
 
 ### eval

results/inference/Llama2-7B/llama2_predict_1_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: llama2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Llama2-7B/llama2_predict_1_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Llama2-7B/llama2_predict_2.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: llama2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Llama2-7B/llama2_predict_2
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Llama2-7B/llama2_predict_2_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: llama2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Llama2-7B/llama2_predict_2_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Llama2-7B/llama2_predict_3.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: llama2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Llama2-7B/llama2_predict_3
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Llama2-7B/llama2_predict_3_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: modelscope/Llama-2-7b-ms
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: llama2
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Llama2-7B/llama2_predict_3_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Qwen-7B/Qwen_predict_1.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: qwen
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Qwen-7B/Qwen_predict_1
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Qwen-7B/Qwen_predict_1_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: qwen
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Qwen-7B/Qwen_predict_1_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Qwen-7B/Qwen_predict_2.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: qwen
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Qwen-7B/Qwen_predict_2
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Qwen-7B/Qwen_predict_2_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: qwen
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Qwen-7B/Qwen_predict_2_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Qwen-7B/Qwen_predict_3.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: qwen
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Qwen-7B/Qwen_predict_3
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

results/inference/Qwen-7B/Qwen_predict_3_single.yaml

@@ -0,0 +1,22 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: qwen
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/inference/Qwen-7B/Qwen_predict_3_single
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: baichuan-inc/baichuan-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: ZhipuAI/chatglm2-6b
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -16,7 +16,7 @@ overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: ./results/lora_sft/Llama2-7B/Llama2-7B_lora_sft_08_15_11_01
+output_dir: ./results/lora_sft/Llama2-7B/llama2_lora_sft_1
 logging_steps: 3
 save_steps: 100
 plot_loss: true
@@ -29,8 +29,9 @@ learning_rate: 1.0e-4
 num_train_epochs: 10.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
-bf16: true
+fp16: true
 ddp_timeout: 180000000
+max_steps: 500
 
 ### eval
 val_size: 0.1

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

@@ -0,0 +1,40 @@
+### 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: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### 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: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### 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: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### 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: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### 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: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

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

@@ -0,0 +1,40 @@
+### model
+model_name_or_path: qwen/Qwen-7B
+
+### 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
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500