This is making me realize that I don’t fully understand the relationship between “instruction-tuned” and “pre-trained”. I thought instruction tuning was a form of fine-tuning, and that fine-tuning comes after the primary training of the model.

I’m not so sure. I think it’s established that the percentage of Republicans who won’t vote for Trump is a lot higher than the percentage that will admit to it in front of their peers.

All we need is a few %, low single digits.