Qwen vl 7b instruct validation accuracy

[achrafs758]

I’m finetunning a qwen-vl-7b-instruct model with lora on an ocr dataset.During the training ,i got around 95% as validation accuracy. I also save a checkpoint every 100 steps. When I load the checkpoint and do inference on the same validation set as during training, the accuracy is much lower. Can someone explain in this context what the accuracy calculated refers to?