This is similar to Llama 2 prompt format, but instead testing whether using different prompt templates negatively affects the outputs of Mixtral 8x7B. They find that the official format that Mistral recommends for its users is the “most censored”, and that “roleplay-oriented presets tend to give better outputs than strictly (bland) assistant-oriented ones”. They also find that “When alignment/censorship is applied during finetuning, it’s closely tied to the prompt format, and deviating from that helps “unleash” the model.
OP aims to address both understanding the effects of prompt formats, but also leveraging to make models more uncensored. “Uncensoring” isn’t as simple as fine-tuning the model. It can be through prompting, changing the llama.cpp —logit-bias flag, etc. More like an art than a systemized tactic.