Since the chat model is local, you can force it out of refusing to respond by giving it a partial answer where it does respond and having it complete that. Haven't tried it myself though.
You could theoretically auto-suppress any tokens that lead to a refusal to answer path and re-generate.
But you’re probably best off grabbing an existing uncensored chat SFT, like the Llama v2 variants trained on Hermes. https://huggingface.co/NousResearch/Nous-Hermes-Llama2-70b