Accuracy reduction of Lora

[yileld]

I used lora trained on qwen2（using the lora code just updated recent weeks）, but the accuracy of test dataset reduced from 75 to 67.
Then I merged lora weights into base model, the accuracy was right.
Is there something wrong with lora? Or some other pipeline difference between lora and normal model?

[Ying1123]

Hi @yileld, would you mind sharing the Lora weights with me so that I can test them? You can join our slack and message me.