y spam?

A spam detection website built with a Svelte front-end and a Flask back-end to serve the prediction model as part of TARUMT - BMCS2123 Natural Language Processing assignment to utilise Meta's Open Pre-trained Transformers to solve real world problems.

To reduce user waiting time, distilbert-base-uncased, a smaller and faster BERT model (available here), has been used to ensure fast tokenisation while having optimal accuracies.

Due to hardware limitations, OPT-350M was used for the final product.

Both Zero-Shot Learning and Few-Shot Learning was applied in this assignment where we used a test sentence: "Congratulation, you have won an iPhone 14. Click this link to claim." to test the accuracy of the model under different conditions.

In our few-shot learning attempt, we had trained OPT-350M on 5574 rows of SMS spam messages (available here).

Results from comparing Zero-Shot / Few Shot Learning

Zero-Shot Learning with 350M parameters

Few-Shot Learning with 350M parameters

Zero-Shot Learning with 1.3B parameters

Screenshots