Just use Llama-2 and not Llama-2-chat. It's not like they are that wildly different in tone. The fine tuning is certainly destructive, but the general tonal b

Just use Llama-2 and not Llama-2-chat. It's not like they are that wildly different in tone. The fine tuning is certainly destructive, but the general tonal biases are typically reflective of the pretrained layers. Imagine what the pie chat makeup of textual data that was used for training would be, look at the normal part of the distribution curve, and that's pretty much exactly what you end up getting with some contextual biasing around prompting. The problem is less that the pretrained models are somehow sinister and more that they create output that is hyper emotional, stubborn, and quick to escalate - much like nearly every thread on social media. Which is contrary to user expectations from AI. A large part of the alignment fine tuning has less to do with safety in terms of keeping a LLM from nuking the world and more to do with preventing a LLM from telling you to go die in a fire after saying its emo poem needed improvement.