LoRA Lab

One tiny model, two personalities. The base model learned plain English, then a LoRA-style adapter (rank 8, under 10% of base params) taught it to talk like a pirate. Same prompt, both answers.