Yep youre right, I don't see an eval on this.

I think you could still essentially have a really cheap Lora-Post training on a frozen base model by doing this. And instead of creating 200 checkpoints for different post-trained directions just adjust a single one on the fly.