support val set in streaming mode

src/llmtuner/tuner/sft/workflow.py

@@ -46,7 +46,7 @@ def run_sft(
         data_collator=data_collator,
         callbacks=callbacks,
         compute_metrics=ComputeMetrics(tokenizer) if training_args.predict_with_generate else None,
-        **split_dataset(dataset, data_args.dev_ratio, training_args.do_train)
+        **split_dataset(dataset, data_args, training_args)
     )
 
     # Keyword arguments for `model.generate`