Strange differences in results when i load saved model to eval

[illusion5556]

First of all, it should be noted that I conducted the experiment on an additional dataset.

When I evaluate on dev during training, the results are P: 0.92545, R: 0.79061, F1: 0.85273.

But after training, I load the saved model, the evaluate results of dev are only P: 0.10000, R: 0.00210, F1: 0.00411.

I barely modified the original code, but I didn't solve the problem.
Looking forward to your reply.