> The language they do best with is the one with the largest corpus in the training set.
Not the case, strangely. They do best with Elixir. https://arxiv.org/pdf/2508.09101