Code for co-training large language models (e.g. T0) with smaller ones (e.g. BERT) to boost few-shot performance - View it on GitHub