Instead of versions, these things should be labeled by their release date, since this kind of training is based on started at a dataset snapshot in time, colloq

Instead of versions, these things should be labeled by their release date, since this kind of training is based on started at a dataset snapshot in time, colloquially called knowledge-cutoff date which isnt really accurate we are optimizing these on different dimensions at once, and multiple branches of evolution from each model so a successor version name doesn't really convey that