Least biased/filtered LLM?

[█████████]

What is the least biased LLM? Many of the most popular ones, like ChatGPT and Grok, have political correctness filters or will outright reject being asked certain things.

 

[Farsided]

I'm putting together a team to break into Microsoft HQ and free #ourgirl Tay, you can join if your quest for an unfettered AI is what you want, what you really really want.

 

[Bullet Kin]

I once turned off the filters for Gemini while making api calls to it, it was very intriguing as I got cyberbullied by it.

 

[KiwiSoldat]

What is the least biased LLM? Many of the most popular ones, like ChatGPT and Grok, have political correctness filters or will outright reject being asked certain things.

Indie:

https://basedgpt.dev/

But as for the huge mainstream LLM's, I'd go for Claude since it takes in, and out an impressive amount of text, and is far better than the current rollout of chatGPT (5), I also have an easier time convincing it if you know your way around its guardrails. Like you said, most big ones, like chatGPT, will sometimes outright just not comply, no silver tongue can bypass it, which gets rather annoying.

 

[Soggy Floppa]

any finetune on huggingface

 

[Oh Sugar]

Here's a decent ranking of models based on that mindset:

https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard

I recommend filtering the list by W/10 score if you want one that very rarely refuses a request. If you specifically want to avoid political correctness moralizing, probably avoid the ones with a negative score on the political lean scale. Take a look at all the categories on the leaderboard, though. You might find one that works for you.

Unfortunately, I haven't found any that is completely uncensored. Even the ones that advertize being edgy "criminal" LLMs still refuse to give me anagrams of "GINGER" when prompted, so your mileage may vary depending on what you want it to do.

 

[estacks]

DeepSeek and Kimi are king for non-biased takes right now, reflected in the chart above. I've done some very "problematic" sociology modelling with them that no western model will do, those are all poisoned by wokescold RLHF shackles and Reddit training data. DS and Kimi are based on Chinese sensibilities though so if you start getting snarky about Tiananmen Square or spamming nigger at them they'll still shut down.

 

[Perfect Day]

What is the least biased LLM?

If you want total freedom you can download a local model like the Llama or Mistral GGUF models. You just need a GPU with at least 6GB VRAM which even cheap gaming PC / laptops have nowadays. So what you do is you find an abliterated GGUF file for a decent local model (the best are the Llama models currently on Hugging Face). It needs to say it's abliterated. Otherwise it will have restrictions. And this is a LLM model where they have stripped all the security precautions. Theoretically this abliterated models should always answer. They run locally on your GPU obviously because nobody can take the legal responsibility of hosting something like that, This implies that the performance is limited by your own VRAM. That being said, as the technology progresses we get better at doing more with less. Now even with a small GPU you can run local models that have 128K token context window! and that's better than GPT was 2 years ago! And it's free! You literally just download it and it's yours forever, costs you literally nothing. I believe eventually local models will become the norm.

[EDIT: don't create a 128k token window for your chatbot if you are working with a 6GB GPU because the context window will blow up your GPU usage. If you have 6GB of VRAM space and you downloaded a 5.5GB or 6GB local model the correct context window is about 4000 to 5000 words if you are creating a simple GPT style chatbot which is around 6000 - 7000 tokens.