[mensetmanusman]

Choice of data to use is alignment. If you train it on the internet, your LLM will spew SSO garbage, so we align it to be more useful.