I am working on a binary prediction classification task, primarily focusing on fine-tuning a BERT model to learn the association between CVEs and CWEs. I’ve structured my task into three phases: first, using a large dataset of CVE and CWE descriptions from 2000 to 2023 for MLM pretraining, which includes 226,572 records; then using the pretrained model weights for fine-tuning on 13,000 CVE records from 2023; and finally, using the fine-tuned model with task-specific weights for inference. For inference, I randomly select 100 pairs of CVEs for the model to predict, only retaining the model weights and setting the model to model.eval(). However, I’ve encountered an issue where, after transferring the pretrained model to fine-tuning, the evaluation metrics significantly improve (accuracy, precision, recall, f1), and both training and validation loss decrease substantially. Yet, when I use the fine-tuned model weights for inference, the accuracy turns out to be worse than before.
What is the reason for this discrepancy?
Here are the hyperparameters used for pretraining and fine-tuning:
Pretrain Model:
batch_size = 16
num_epochs = 10
learning_rate = 2e-5
eps = 1e-8
beta1 = 0.9
beta2 = 0.99
weight_decay = 0.01
total_steps = num_epochs * len(train_loader)
warmup_steps = total_steps // 10
early_stopping_patience = 2
mask_prob = 0.15
replace_mask_prob = 0.8
random_replace_prob = 0.10
keep_original_prob = 0.10
Train Model:
learning_rate = 2e-5
batch_size = 16
epoch = 10
I hope someone can help address this issue. If more detailed code or data is needed, I am willing to provide it. Thank you.