fix padding subword labeling

iwslt2011.py

@@ -269,6 +269,7 @@ class IWSLT11(datasets.GeneratorBasedBuilder):
                     raise ValueError('tokenizer argument has to be passed to load_dataset')
                 else:
                     words = tokens[i].words
+                    input_ids = tokens['input_ids'][i]
                     offsets = np.array(tokens['offset_mapping'][i])
                     enc_labels = np.array([self.config.label_subword.name]*len(offsets), dtype=object)
                     count = 0
@@ -276,10 +277,12 @@ class IWSLT11(datasets.GeneratorBasedBuilder):
                         if word_id is not None and (j == 0 or words[j-1] != word_id):
                             enc_labels[j] = labels[count].name
                             count += 1
+                        elif input_ids[j] == self.config.tokenizer.pad_token_id:
+                            enc_labels[j] = LabelSubword.IGNORE.name
                     labels = enc_labels
                 yield i, {
                             "ids": ids,
-                            "tokens": tokens['input_ids'][i],
+                            "tokens": input_ids,
                             "labels": labels,
                         }
             logging.info(f"Loaded number of tokens = {token_len}")