Pytorch (RE)-Implementation of Grokking Phenomenon

This is a pytorch re-implementation of Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets.

I thought this would be a good paper to reproduce since this would allow me to code and train a GPT style model from scratch.

References used for the Code :-

1. MinGPT by Karpathy

Accuracy Loss Curves for Adam (with any weight decay)

Accuracy Loss Curves for AdamW ( λ = 1 )