diff --git a/cbert_utils.py b/cbert_utils.py
index b441d4b..a85d313 100755
--- a/cbert_utils.py
+++ b/cbert_utils.py
@@ -197,7 +197,7 @@ def extract_features(tokens_a, tokens_label, max_seq_length, tokenizer):
     masked_lm_probs = 0.15
     max_predictions_per_seq = 20
     rng = random.Random(12345)
-    original_masked_lm_labels = [-1] * max_seq_length
+    original_masked_lm_labels = [-100] * max_seq_length
     (output_tokens, masked_lm_positions, 
     masked_lm_labels) = create_masked_lm_predictions(
             tokens, masked_lm_probs, original_masked_lm_labels, max_predictions_per_seq, rng, tokenizer)
@@ -271,4 +271,4 @@ def create_masked_lm_predictions(tokens, masked_lm_probs, masked_lm_labels,
         masked_lm_labels[index] = convert_tokens_to_ids([tokens[index]], tokenizer)[0]
         output_tokens[index] = masked_token
         masked_lm_positions.append(index)
-    return output_tokens, masked_lm_positions, masked_lm_labels
\ No newline at end of file
+    return output_tokens, masked_lm_positions, masked_lm_labels