chore: _

config.py

@@ -8,8 +8,8 @@ class DataConfig:
 ---
 Translation:
 {}"""
-    train_split: float = 0.9
-    max_samples: int | None = 3000
+    train_split: float = 0.95
+    max_samples: int | None = 5000
 
 
 @dataclass
@@ -31,7 +31,7 @@ class TrainingConfig:
     base_model: str = "unsloth/Qwen2.5-7B"
     max_seq_length: int = 6144
     dtype: str | None = None
-    load_in_4bit: bool = False
+    load_in_4bit: bool = True
 
     # LoRA
     lora_r: int = 64
@@ -48,7 +48,7 @@ class TrainingConfig:
             "down_proj",
         ]
     )
-    use_gradient_checkpointing: str = "unsloth"
+    use_gradient_checkpointing: str = True
     random_state: int = 3407
     use_rslora: bool = False
     loftq_config: dict | None = None
@@ -56,11 +56,11 @@ class TrainingConfig:
     # training args
     per_device_train_batch_size: int = 16
     gradient_accumulation_steps: int = 2
-    warmup_ratio: float = 0.05
+    warmup_ratio: float = 0.1
     max_grad_norm: float = 1.0
     num_train_epochs: float = 1
     learning_rate: float = 5e-4
-    weight_decay: float = 0
+    weight_decay: float = 0.01
     lr_scheduler_type: str = "cosine"
     logging_steps: int = 1
 
@@ -70,15 +70,15 @@ class TrainingConfig:
     save_total_limit: int | None = 3
 
     # dataset
-    dataset_num_proc: int = 8
+    dataset_num_proc: int = 4
     packing: bool = True
 
     # eval
     fp16_full_eval: bool = True
-    per_device_eval_batch_size: int = 16
-    eval_accumulation_steps: int = 4
+    per_device_eval_batch_size: int = 64
+    eval_accumulation_steps: int = 1
     eval_strategy: str = "steps"
-    eval_steps: int = 5
+    eval_steps: int = 10
 
     # output
     output_dir: str = "/workspace/output/"