Finetune the base language model (llama2 7B) with (output, instruction) pairs {(yi, xi)} from the seed data to obtain a backward model Myx := p(x|y). In other words, finetune a model that uses the output to predict the instruction. Use the openassistant-guanaco training set dataset