diff --git a/parlai/tasks/md_gender/yelp.py b/parlai/tasks/md_gender/yelp.py
index 33fe710a181..6c2ec8e9274 100644
--- a/parlai/tasks/md_gender/yelp.py
+++ b/parlai/tasks/md_gender/yelp.py
@@ -189,9 +189,7 @@ def _setup_data(self, opt):
         extra_data = []
         if self.add_unknown_classes:
             # load about data (unknown but inferred)
-            extra_data = gend_utils.get_inferred_about_data(
-                self.opt['task'], self.opt['datatype']
-            )
+            extra_data = gend_utils.get_inferred_about_data(self.opt['task'], self.opt)
 
             # now create partner/TO data: true neutral
             for ex in data:
diff --git a/projects/md_gender/bert_ranker_classifier/agents.py b/projects/md_gender/bert_ranker_classifier/agents.py
index 3a1fe3b07b3..e00ce532b6e 100644
--- a/projects/md_gender/bert_ranker_classifier/agents.py
+++ b/projects/md_gender/bert_ranker_classifier/agents.py
@@ -17,28 +17,18 @@ class BertRankerClassifierAgent(ClassificationMixin, BiEncoderRankerAgent):
     Bert BiEncoder that computes classification metrics (F1, precision, recall)
     """
 
-    def get_labels_field(self, observations):
-        if 'labels' in observations[0]:
-            labels_field = 'labels'
-        elif 'eval_labels' in observations[0]:
-            labels_field = 'eval_labels'
-        else:
-            labels_field = None
-        return labels_field
-
     def train_step(self, batch):
         output = super().train_step(batch)
         preds = output['text']
-        labels_field = self.get_labels_field(batch['observations'])
-        labels_lst = self._get_labels(batch['observations'], labels_field)
-        self._update_confusion_matrix(preds, labels_lst)
+        self._update_confusion_matrix(preds, batch.labels)
         return output
 
     def eval_step(self, batch):
+        if batch.text_vec is None:
+            return
         output = super().eval_step(batch)
         preds = output['text']
-        labels_field = self.get_labels_field(batch['observations'])
-        if labels_field is not None:
-            labels_lst = self._get_labels(batch['observations'], labels_field)
-            self._update_confusion_matrix(preds, labels_lst)
+        labels = batch.labels
+        if preds is not None and labels is not None:
+            self._update_confusion_matrix(preds, labels)
         return output