Various efficient attention encoder-style architectures distilled into student models with half the hidden layers, plus a long-context NER dataset