Yeah, this is probably true. In training, I aimed to allow it to provide slightly broader alternatives, but not more specific alternatives, like this one is.
Since all groupthink is a form of consensus, under the rules I’ve been following it would be acceptable for it to highlight “groupthink” and provide “consensus” as an alternative, but not the other way around.
Seeing how competent these newer models are was quite helpful, and I have started using ChatGPT o4-mini-high to generate training sets for a new finetuned model. It does quite well, and once I get the hang of this, I should be able to distill the results and improve the performance of my model much more quickly. Wish me luck!