Prototype after contrastive learning first pass. Unlikely to be good due to data duplication

Should be a base model with no MLM head