5 ms·
Yes. Meta deliberately let their responsible AI people heavily censor the chat-tuned model. Meta also released the uncensored base model, on which the open sou
by throwaway9274 3y ago
Yes. Meta deliberately let their responsible AI people heavily censor the chat-tuned model.
Meta also released the uncensored base model, on which the open source community then performed its own chat fine tunes. This was a canny strategy to avoid negative press.
Mistral saw Meta’s approach, and instead chose to deliberately court the negative press, because attention is more valuable to them as a startup than opprobrium is damaging.
- astrange 3y agoSince 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.
- throwaway9274 3y agoDefinitely 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 https://huggingface.co/NousResearch/Nous-Hermes-Llama2-70b