chore: _

config.py

@@ -9,7 +9,7 @@ class DataConfig:
 Translation:
 {}"""
     train_split: float = 0.95
-    max_samples: int | None = 5000
+    max_samples: int | None = None
 
 
 @dataclass
@@ -31,11 +31,11 @@ class TrainingConfig:
     base_model: str = "unsloth/Qwen2.5-7B"
     max_seq_length: int = 6144
     dtype: str | None = None
-    load_in_4bit: bool = True
+    load_in_4bit: bool = False
 
     # LoRA
-    lora_r: int = 64
-    lora_alpha: int = 128
+    lora_r: int = 16
+    lora_alpha: int = 32
     lora_dropout: float = 0
     target_modules: list[str] = field(
         default_factory=lambda: [
@@ -49,18 +49,19 @@ class TrainingConfig:
         ]
     )
     use_gradient_checkpointing: str = "unsloth"
-    random_state: int = 3407
+    random_state: int = 42
     use_rslora: bool = False
     loftq_config: dict | None = None
 
     # training args
     per_device_train_batch_size: int = 16
-    gradient_accumulation_steps: int = 2
-    warmup_ratio: float = 0.1
+    gradient_accumulation_steps: int = 4
+    # warmup_ratio: float = 0.1
+    warmup_steps: int = 80
     max_grad_norm: float = 1.0
-    num_train_epochs: float = 1
-    learning_rate: float = 5e-4
-    weight_decay: float = 0.01
+    num_train_epochs: float = 3
+    learning_rate: float = 1e-5
+    weight_decay: float = 0.001
     lr_scheduler_type: str = "cosine"
     logging_steps: int = 1
 
@@ -70,15 +71,15 @@ class TrainingConfig:
     save_total_limit: int | None = 3
 
     # dataset
-    dataset_num_proc: int = 4
+    dataset_num_proc: int = 8
     packing: bool = True
 
     # eval
     fp16_full_eval: bool = True
-    per_device_eval_batch_size: int = 8
-    eval_accumulation_steps: int = 2
+    per_device_eval_batch_size: int = 16
+    eval_accumulation_steps: int = 1
     eval_strategy: str = "steps"
-    eval_steps: int = 10
+    eval_steps: int = 100
 
     # output
     output_dir: str = "/workspace/output/"