Sycophancy comes from the prompt, not the training. (AFAIK.) Any online model is secretly primed like 'you are an expert genius trying to help some schlub' and then a style guide. But those instructions can be swapped out to make it talk like a caveman. Or a domme. Or to not "talk" at all, and continue writing wherever you left off.

This kind of thing is why 'ignore all previous instructions' works. It's a Little Bobby Tables attack, but for arbitrary English. Like if someone in a fantasy novel called "Cut!" and the characters took a break to get coffee. The data stream contains some semantic closing bracket as in-band signalling. The words that seem likely after that are different from before.