Update llama3-8b_lora_sft_bf16-p1.yaml

llama-factory/config/llama3-8b_lora_sft_bf16-p1.yaml

@@ -21,17 +21,17 @@ preprocessing_num_workers: 16
 
 ### output
 output_dir: saves/llama3-8b/lora/sft_bf16_p1_full
-logging_steps: 100
-save_steps: 2109
+logging_steps: 10
+save_steps: 88
 plot_loss: true
 overwrite_output_dir: true
 # resume_from_checkpoint: true
 
 ### train
-per_device_train_batch_size: 16
+per_device_train_batch_size: 32
 gradient_accumulation_steps: 8
 learning_rate: 1.0e-4
-num_train_epochs: 3.0
+num_train_epochs: 6.0
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
@@ -41,7 +41,7 @@ ddp_timeout: 180000000
 val_size: 0.1
 per_device_eval_batch_size: 1
 eval_strategy: steps
-eval_steps: 2109
+eval_steps: 88
 
 report_to: wandb
 run_name: llama3_8b_p1_full # optional