News for hackers. Stuff that matters.
Re:French AI startup Mistral releases ‘unmoderated’ chatbot ... by throwaway9274 on Saturday September 30, 2023 @08:22PM (#37721291)
Definitely one of the benefits of local models.

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

There are no comments on this comment yet.