Refactor SFTTrainer configuration in train.py to remove data_collator from the SFT config, preventing duplication and enhancing clarity in trainer setup.

train.py

@@ -200,6 +200,10 @@ def create_trainer(
             **cfg.training.sft.data_collator,
         )
 
+        # Create SFT config without data_collator to avoid duplication
+        sft_config = OmegaConf.to_container(cfg.training.sft, resolve=True)
+        sft_config.pop('data_collator', None)  # Remove data_collator from config
+
         trainer = SFTTrainer(
             model=model,
             tokenizer=tokenizer,
@@ -207,7 +211,7 @@ def create_trainer(
             eval_dataset=dataset["validation"],
             args=training_args,
             data_collator=data_collator,
-            **cfg.training.sft,
+            **sft_config,
         )
         logger.info("Trainer created successfully")
         return trainer