diff --git a/mantis/train/data.py b/mantis/train/data.py
index 64912e4..51b6d6f 100644
--- a/mantis/train/data.py
+++ b/mantis/train/data.py
@@ -1396,7 +1396,7 @@ def _right_pad_inputs_with_attention_mask(self, model_inputs: List[Dict]):
                 if k == 'input_ids':
                     # add padding
                     max_length = max([inputs[k].shape[1] for inputs in model_inputs])
-                    pad_token_id = self.tokenizer.pad_token_id
+                    pad_token_id = self.processor.tokenizer.pad_token_id
                     # pad all inputs to the same length
                     results[k] = torch.cat(
                         [