train: change yaml

examples/train_lora/llama2_lora_sft.yaml

@@ -13,10 +13,9 @@ template: llama2
 cutoff_len: 1024
 max_samples: 100000
 overwrite_cache: true
-preprocessing_num_workers: 16
 
 ### output
-output_dir: ./saves/LLaMA2-7B/lora/train_24_8_13_08_13
+output_dir: ./saves/LLaMA2-7B/lora/train_24_8_13_10_02
 logging_steps: 3
 save_steps: 100
 plot_loss: true
@@ -25,16 +24,15 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 2
 gradient_accumulation_steps: 8
-learning_rate: 1.5e-4
+learning_rate: 5.0e-05
 num_train_epochs: 10
 lr_scheduler_type: cosine
-warmup_ratio: 0.1
 fp16: true
-ddp_timeout: 180000000
-
-### eval
-val_size: 0.1
-per_device_eval_batch_size: 1 # 1 最大
-eval_strategy: steps
-eval_steps: 500
 
+max_grad_norm: 1.0
+warmup_steps: 0
+optim: adamw_torch
+report_to: none
+lora_rank: 8
+lora_alpha: 16
+lora_dropout: 0.1