Take the phi$-2$ LLM example. Adjust train$\_$test$\_$split function to generate approximately $400$ positive and $400$ negative sentiment feature vectors. Similarly, generate $50$ positive and $50$ negative test samples. Using a batch$-$size of $4$:

a$)$ Build logistic regression classifier. Report your accuracy on train and test sets. Explain if there are any overfit or underfit observed?

b$)$ Build a neural network with two hidden layers. Use $5$ nodes in each layer. Use

RELU nonlinearity. Try three different learning rates. Explain if there are any overfit or underfit observed?

c$)$ Build an encoder$-$only transformer network. Use two heads in attention. Try two

different learning rates and number of layers. Explain if there are any overfit or underfit observed?

d$)$ Do the part$-$c but remove residual connections and layer normalization.