“It does not have any moderation mechanism.”
Mistral 7B ships under Apache 2.0 and claims to beat Llama 2 13B on every benchmark Mistral chose to run, through its own evaluation pipeline. The instruct version was tuned on public Hugging Face datasets with no proprietary data. It has no guardrails at all, and Mistral says it looks forward to talking with the community about adding some. The safety plan is a conversation scheduled for after the weights are already out.