Add DataCollatorForLanguageModeling to trainer configuration in train.py for improved data handling during training.

train.py

@@ -34,6 +34,7 @@ from datasets import (
 )
 from transformers import AutoTokenizer, Trainer, TrainingArguments
 from trl import SFTTrainer
+from trl.data.data_collator import DataCollatorForLanguageModeling
 
 # Configuration
 max_seq_length = 2048  # Auto supports RoPE Scaling internally
@@ -211,8 +212,8 @@ def create_trainer(
                 logging_steps=1,
                 save_strategy="steps",
                 save_steps=30,
-                eval_strategy="steps",  # Match save_strategy
-                eval_steps=30,  # Match save_steps
+                eval_strategy="steps",
+                eval_steps=30,
                 save_total_limit=2,
                 optim="adamw_8bit",
                 weight_decay=0.01,
@@ -224,6 +225,11 @@ def create_trainer(
                 metric_for_best_model="eval_loss",
                 greater_is_better=False,
             ),
+            data_collator=DataCollatorForLanguageModeling(
+                tokenizer=tokenizer,
+                mlm=False,
+                pad_to_multiple_of=8,
+            ),
         )
         logger.info("Trainer created successfully")
         return trainer