> the DeepMind team built a custom dataset from CodeContests from two previous datasets, with over 13,500 challenges. Each came with an explanation of the task at hand, and multiple potential solutions across multiple languages. The result is a massive library of training data tailored to the challenge at hand.
Isn't this just over-fitting the model?