diff --git a/src/lerobot/datasets/dataset_reader.py b/src/lerobot/datasets/dataset_reader.py
index 3720a5084..dd496142f 100644
--- a/src/lerobot/datasets/dataset_reader.py
+++ b/src/lerobot/datasets/dataset_reader.py
@@ -87,7 +87,7 @@ class DatasetReader:
         """Attempt to load from local cache. Returns True if data is sufficient."""
         try:
             self.hf_dataset = self._load_hf_dataset()
-        except (FileNotFoundError, NotADirectoryError):
+        except (FileNotFoundError, NotADirectoryError, ValueError):
             self.hf_dataset = None
             return False
         if not self._check_cached_episodes_sufficient():
diff --git a/src/lerobot/datasets/io_utils.py b/src/lerobot/datasets/io_utils.py
index cee6cfba8..2d1b5433d 100644
--- a/src/lerobot/datasets/io_utils.py
+++ b/src/lerobot/datasets/io_utils.py
@@ -78,7 +78,10 @@ def load_nested_dataset(
     with SuppressProgressBars():
         # We use .from_parquet() memory-mapped loading for efficiency
         filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
-        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+        try:
+            return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+        except ValueError:
+            raise ValueError(f"Failed to load parquet files in {pq_dir}, make sure the dataset is valid and is not missing any files.")
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int: