diff --git a/trankit/adapter_transformers/tokenization_utils.py b/trankit/adapter_transformers/tokenization_utils.py
index 3bb91c0..663ddde 100644
--- a/trankit/adapter_transformers/tokenization_utils.py
+++ b/trankit/adapter_transformers/tokenization_utils.py
@@ -924,7 +924,15 @@ def _from_pretrained(cls, pretrained_model_name_or_path, *init_inputs, **kwargs)
         if pretrained_model_name_or_path in s3_models:
             # Get the vocabulary from AWS S3 bucket
             for file_id, map_list in cls.pretrained_vocab_files_map.items():
-                vocab_files[file_id] = map_list[pretrained_model_name_or_path]
+                basename_from_s3 = os.path.basename(map_list[pretrained_model_name_or_path])
+                model_prefix  = slice(len(pretrained_model_name_or_path) + 1, None)
+                fname = basename_from_s3 [model_prefix]  # filename as mentioned on hf.co/models
+                model_folder_exists = os.path.isdir(pretrained_model_name_or_path) 
+                fname_exists = os.path.exists(os.path.join(pretrained_model_name_or_path, fname))
+                if model_folder_exists and fname_exists:
+                    vocab_files[file_id] = os.path.join(pretrained_model_name_or_path, fname)
+                else:
+                    vocab_files[file_id] = map_list[pretrained_model_name_or_path]
             if (
                 cls.pretrained_init_configuration
                 and pretrained_model_name_or_path in cls.pretrained_init_configuration