I'm in the process of updating the pytorch benchmarks repo to use newer torchtext APIs. These benchmarks use an adapted version of this example notebook. If it makes sense, I'm happy to submit a PR to similarly update the examples here (and in other notebooks where relevant).