Final results' huge gap while using & not using pretrained weights?

[tonylin52]

I train the model without pretrained weights, the final results are as below [according to the paper, ser f1 should be 83.39%, re f1 should be 74.91%]: However, when I train the model with pretrained weight, the results look much better. But decrease after several steps:

[tonylin52]

In my exprience, pretrained process could improve around 30%-50% preformence. But in this case: 100%-200%.

[ccx1997]

I train the model without pretrained weights, the final results are as below [according to the paper, ser f1 should be 83.39%, re f1 should be 74.91%]

In the first row in Tab. 3, it means the model is only pretrained using MVLM.

Such a large backbone initialized from scratch is prone to get overfitting on a small dataset.

[tonylin52]

I train the model without pretrained weights, the final results are as below [according to the paper, ser f1 should be 83.39%, re f1 should be 74.91%]

In the first row in Tab. 3, it means the model is only pretrained using MVLM.

Such a large backbone initialized from scratch is prone to get overfitting on a small dataset.

Thanks! But I tried evaluation using training dataset, it seems not overfitting for the relation extraction task while training without pretrained weights: