fix: update dataclass objects directly instead of creating new variab…

…les (#418) Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com>
foundation-model-stack · Dec 13, 2024 · a89f76b · a89f76b
1 parent a098e70
commit a89f76b
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/tuning/sft_trainer.py b/tuning/sft_trainer.py
@@ -295,9 +295,9 @@ def train(
     (
         formatted_train_dataset,
         formatted_validation_dataset,
-        dataset_text_field,
+        data_args.dataset_text_field,
         data_collator,
-        max_seq_length,
+        train_args.max_seq_length,
         dataset_kwargs,
     ) = process_dataargs(data_args, tokenizer, train_args, additional_data_handlers)
     additional_metrics["data_preprocessing_time"] = (
@@ -327,7 +327,7 @@ def train(
     }
 
     additional_args = {
-        "dataset_text_field": dataset_text_field,
+        "dataset_text_field": data_args.dataset_text_field,
         "dataset_kwargs": dataset_kwargs,
     }
     training_args = SFTConfig(**transformer_kwargs, **additional_args)