diff --git a/src/llmtuner/train/orpo/trainer.py b/src/llmtuner/train/orpo/trainer.py
index 50b999f8..f5b7ff42 100644
--- a/src/llmtuner/train/orpo/trainer.py
+++ b/src/llmtuner/train/orpo/trainer.py
@@ -99,7 +99,7 @@ class CustomORPOTrainer(DPOTrainer):
         """
         metrics = {}
         chosen_logps, rejected_logps, chosen_logits, rejected_logits = self.concatenated_forward(model, batch)
-        sft_loss = chosen_logps
+        sft_loss = -chosen_logps
         odds_ratio_loss = self.odds_ratio_loss(chosen_logps, rejected_logps)
         batch_loss = (sft_loss + self.beta * odds_ratio_loss).mean()