[parser] support omegaconf (#7793)

src/llamafactory/train/dpo/workflow.py

@@ -63,9 +63,6 @@ def run_dpo(
     else:
         ref_model = None
 
-    # Update arguments
-    training_args.remove_unused_columns = False  # important for multimodal and pairwise dataset
-
     # Initialize our Trainer
     trainer = CustomDPOTrainer(
         model=model,

src/llamafactory/train/kto/workflow.py

@@ -59,9 +59,6 @@ def run_kto(
     else:
         ref_model = create_ref_model(model_args, finetuning_args)
 
-    # Update arguments
-    training_args.remove_unused_columns = False  # important for multimodal and pairwise dataset
-
     # Initialize our Trainer
     trainer = CustomKTOTrainer(
         model=model,

src/llamafactory/train/rm/workflow.py

@@ -48,9 +48,6 @@ def run_rm(
         template=template, model=model, pad_to_multiple_of=8, **tokenizer_module
     )
 
-    # Update arguments
-    training_args.remove_unused_columns = False  # important for multimodal and pairwise dataset
-
     # Initialize our Trainer
     trainer = PairwiseTrainer(
         model=model,

src/llamafactory/train/sft/workflow.py

@@ -65,11 +65,6 @@ def run_sft(
         **tokenizer_module,
     )
 
-    # Override the decoding parameters of Seq2SeqTrainer
-    training_args.generation_max_length = training_args.generation_max_length or data_args.cutoff_len
-    training_args.generation_num_beams = data_args.eval_num_beams or training_args.generation_num_beams
-    training_args.remove_unused_columns = False  # important for multimodal dataset
-
     # Metric utils
     metric_module = {}
     if training_args.predict_with_generate: