Well, haven't we seen similar results before? IIRC finetuning for safety or "alignment" degrades the model too. I wonder if it is true that finetuning a model for anything will make it worse. Maybe simply because there is just orders of magnitudes less data available for finetuning, compared to pre-training.