Comment by zamalek

I don't get it, people are going to train or tune models on uncensored data regardless of what the original researchers do. Uncensored models are already readily available for Llama, and significantly outperform censored models of a similar size.

Output sanitization makes sense, though.