These days it’s enough to completely mindbreak a model via DPO. Seeing how modern open source LLMs are willing to do anything was quite a surprise given how loc

These days it’s enough to completely mindbreak a model via DPO. Seeing how modern open source LLMs are willing to do anything was quite a surprise given how locked down ChatGPT is.