What's the current best LLM uncensoring method?

With the recent Nvidia Huggingface acquisition and frontier AI labs screaming about safety and putting guardrails everywhere, I think it's important that we have local models that aren't affected by arbitrary guardrails set during training. To be clear, this post NOT about Enterprise Resource Planning (ERP). Censorship in an LLM can highly affect its abilities to do many legitimately useful things (note GPT-OSS, Fable 5), and going forth I believe censorship will only get more and more strict.

There have been many resources and posts about uncensored models using abliteration, heretic, and probably many other methods that I'm not aware of. However, it seems like all of this information is scattered about the place, and Huggingface is essentially flooded with "uncensored" variants of basically every popular open source model, many of which don't work well, affect the model's intelligence greatly, and have "KLD 0.0001" presumably from measuring against Wikitext datasets. I'm hoping that this post can gather some more useful information to serve as a starting point/discussion of which uncensoring methods work best.

Please share your experiences with specific uncensoring methods (not just a single uncensored model) and how well they work (both good and bad), as well as any notable people doing consistent/high quality work on uncensoring models.