Finetune on My Own Pre-trained Model

[SaikaOfficial]

Fine-tuning using the provided pre-trained model was able to achieve good, paper-like results, but using the pre-trained code to train the model itself and then fine-tuning gave very poor results.Did something go wrong?
picture below is fine-tune and test result on my own pre-trained model

but, this is on the proveied model.

I followed the steps exactly and was able to see my training process on the tensorboard, and there didn't seem to be any errors.