[NeuralChat] use validation dataset for evaluation.

intel_extension_for_transformers/llm/finetuning/finetuning.py

@@ -445,12 +445,10 @@ def concatenate_data(dataset, max_seq_length):
             )
 
         if training_args.do_eval:
-            if "test" not in tokenized_datasets:
-                self.logger.info('Splitting train dataset in train and validation according to `eval_dataset_size`')
-                tokenized_datasets = tokenized_datasets["train"].train_test_split(
-                    test_size=data_args.eval_dataset_size, shuffle=True, seed=42
-                )
-            eval_dataset = tokenized_datasets["test"]
+            if "validation" not in tokenized_datasets:
+                raise ValueError("--do_eval requires a validation dataset")
+
+            eval_dataset = tokenized_datasets["validation"]
             if data_args.max_eval_samples is not None:
                 eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
 

intel_extension_for_transformers/neural_chat/config.py

@@ -228,9 +228,6 @@ class DataArguments:
             )
         },
     )
-    eval_dataset_size: int = field(
-        default=500, metadata={"help": "Size of validation dataset."}
-    )
     streaming: bool = field(default=False, metadata={"help": "Enable streaming mode"})
     preprocessing_num_workers: Optional[int] = field(
         default=None,