fix text-only dataset

[tomiock]

• these datasets should have separate cookers and use the same task encoder as the regular vision datasets that we are already training on