shawhin/bert-phishing-classifier_student

bert-phishing-classifier_student

This model is modified version of distilbert/distilbert-base-uncased trained via knowledge distillation from shawhin/bert-phishing-classifier_teacher using the shawhin/phishing-site-classification dataset. It achieves the following results on the testing set:

• —Loss (training): 0.0563
• —Accuracy: 0.9022
• —Precision: 0.9426
• —Recall: 0.8603
• —F1 Score: 0.8995

Model description

Student model for knowledge distillation example.

Video | Blog | Example code

Intended uses & limitations

This model was created for educational purposes.

Training and evaluation data

The Training, Testing, and Validation data are available here: shawhin/phishing-site-classification.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 0.0001
• —trainbatchsize: 32
• —evalbatchsize: 32
• —num_epochs: 5
• —temperature: 2.0
• —adam optimizer alpha: 0.5