update train script

train_with_unsloth.py

@@ -142,19 +142,19 @@ def clean_assistant_marker(example):
 # clean: <|im_start|>assistant\n\n -> <|im_start|>assistant\n
 dataset = dataset.map(clean_assistant_marker, batched=False)
 
-new_dataset = dataset.train_test_split(test_size=0.01)
+new_dataset = dataset.train_test_split(test_size=0.1)
 
 # Configure training arguments
 training_args = SFTConfig(
     fp16_full_eval=False,
     per_device_train_batch_size=1,
-    gradient_accumulation_steps=2,
+    gradient_accumulation_steps=1,
     per_device_eval_batch_size=1,
-    eval_accumulation_steps=4,
+    eval_accumulation_steps=1,
     evaluation_strategy="steps",
-    eval_steps=100,
+    eval_steps=1000,
     save_strategy="steps",
-    save_steps=500,
+    save_steps=1000,
     load_best_model_at_end=True,
     metric_for_best_model="eval_loss",
     greater_is_better=False,