Fix path to metadata.csv

dataset.py

@@ -93,9 +93,10 @@ class JeliASR(datasets.GeneratorBasedBuilder):
         - TRAIN: file paths containing "/train/"
         - TEST: file paths containing "/test/"
         """
+        base_dir = self.config.data_dir or ""
         # If someone calls load_dataset("RobotsMali/jeli-asr", data_dir="jeli-asr"),
         # self.config.data_dir will be set to "jeli-asr" inside the downloaded repo.
-        metadata_path = os.path.join(dl_manager.manual_dir or "", "metadata.csv")
+        metadata_path = os.path.join(base_dir, "metadata.csv")
 
         # We do not need to call dl_manager.download_and_extract here because
         # the audio + CSV are already on the Hub. The Datasets library will handle