Safety is just the latest trojan horse being used by big tech to try and control how people use their computers. I definitely belive in responsible use of AI, but I don't belive that any of these companies have my best interests at heart and that I should let them tell me what I can do with a computer.

Those who trade liberty for security get neither and all that.

Actually, leaving out whether “safety” is inherently “bullshit” [0], it is both, Llama Guard is a model, serving a similar function to the OpenAI moderation API, but in a weights-available model.

[0] “AI safety”, is often, and the movement that popularized the term is entirely, bullshit and largely a distraction from real and present social harms from AI. OTOH, relatively open tools that provide information to people building and deploying LLMs to understand their capacities in sensitive areas and the actual input and output are exactly the kind of things people who want to see less centralized black-box heavily censored models and more open-ish and uncensored models as the focus of development should like, because those are the things that make it possible for institutions to deploy such models in real world, significant applications.

The safety here is not just "don't mention potentially controversial topics".

The safety here can also be LLMs working within acceptable bounds for the usecase.

Let's say you had a healthcare LLM that can help a patient navigate a healthcare facility, provide patient education, and help patients perform routine administrative tasks at a hospital.

You wouldn't want the patient to start asking the bot for prescription advice and the bot to come back with recommending dosages change, or recommend a OTC drug with adverse reactions to their existing prescriptions, without a provider reviewing that.

We know that currently many LLMs can be prompted to return nonsense very authoritatively, or can return back what the user wants it to say. There's many settings where that is an actual safety issue.

Well it is a new model, it's just a safety bullshit model (your words).

But the datasets could be useful in their own right. I would consider using the codesec one as extra training data for a code-specific LLM – if you're generating code, might as well think about potential security implications.