To get pure grokking, you need a model large enough to easily memorize the entire training data and keep training for a long time after memorization. In practice, you'll probably use a more realistically-sized model that might grok on some subset of the data, but not so strongly that it's extremely obvious.