diff --git a/examples/adversarial/utils_hans.py b/examples/adversarial/utils_hans.py
index 8f230fad981..1860c98dccf 100644
--- a/examples/adversarial/utils_hans.py
+++ b/examples/adversarial/utils_hans.py
@@ -255,7 +255,11 @@ class HansProcessor(DataProcessor):
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "heuristics_evaluation_set.txt")), "dev")
 
     def get_labels(self):
-        """See base class."""
+        """See base class.
+        Note that we follow the standard three labels for MNLI
+        (see :class:`~transformers.data.processors.utils.MnliProcessor`)
+        but the HANS evaluation groups `contradiction` and `neutral` into `non-entailment` (label 0) while
+        `entailment` is label 1."""
         return ["contradiction", "entailment", "neutral"]
 
     def _create_examples(self, lines, set_type):
@@ -268,7 +272,7 @@ class HansProcessor(DataProcessor):
             text_a = line[5]
             text_b = line[6]
             pairID = line[7][2:] if line[7].startswith("ex") else line[7]
-            label = line[-1]
+            label = line[0]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label, pairID=pairID))
         return examples