LaMDA consists of several fine-tunings of a single pre-trained model. Each of them generate candidate responses, the candidate responses are scored (using yet a

LaMDA consists of several fine-tunings of a single pre-trained model. Each of them generate candidate responses, the candidate responses are scored (using yet another fine-tuning of the model, iirc) and the best is selected. One expects that the version of LaMDA for internal and raters use does not have to be frugal.