Fix default eval split to 'test'

The iconclass-vlm-sft dataset has 'train' and 'test' splits, not 'valid'

iconclass-vlm-sft.py

@@ -262,7 +262,7 @@ def main(
     max_seq_length: int = 2048,
     hf_token: str = None,
     dataset_split: str = "train",
-    eval_split: str = "valid",
+    eval_split: str = "test",
     private: bool = False,
     push_to_hub: bool = True,
 ):
@@ -603,8 +603,8 @@ Examples:
     )
     dataset_group.add_argument(
         "--eval-split",
-        default="valid",
-        help="Dataset split to use for evaluation (default: valid)",
+        default="test",
+        help="Dataset split to use for evaluation (default: test)",
     )
     dataset_group.add_argument(
         "--max-seq-length",