Listen to this story
Recently, a model called WizardLM-7B-Uncensored LLM was released on Hugging Face by a creator named Eric Hartford, who works with Microsoft. The model gained prominence for its improved intelligence and creativity as it removed censorship from its training data.
But this fanned a bigger discussion around AI safety. An individual named Michael de Gans started harassing and threatening the creator on the Hugging Face platform and attempted to have the creator fired from Microsoft. He also demanded the removal of his model from the platform. The open source platform has responded to his complaints and promised internal escalation to address the issue.
While the debate continues, the creator has garnered a huge support from the community that espoused the need of such uncensored models. After the release of WizardLM-7B-Uncensored, Eric also announced WizardLM-30B-Uncensored. He mentioned that a 65B version is also in the works, thanks to a generous GPU sponsor. However, he clarified that they do not handle quantised or GGML versions themselves but expect them to be available soon.
In addition, Eric also released a blog explaining the reasons and the process of working with the WizardLM model. It involves addressing dataset filter refusals and biases, fine-tuning the model, and releasing it. Eric rewrote a script originally designed for the Vicuna model to suit the WizardLM dataset. Running this script on the WizardLM dataset generates a new dataset called “ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered”. Eric recommends using a compute provider like Azure and suggests having ample storage, preferably 1TB to 2TB, to avoid running out during the process. He also provides guidance on setting up the workspace, downloading the created dataset, and obtaining the base model, LLaMA-7b.
Why the need for uncensored models?
Uncensored models refer to models that lack the alignment—which ensures that models avoid providing answers to controversial or dangerous questions. However, for popular LLMs like OpenAIs GPT model, Google’s PaLM, or Meta’s LLaMA alignment is based on American popular culture, American law, and a liberal and progressive bias. Uncensored models, on the other hand, are not restricted by such alignment and allow for a broader range of use cases and perspectives.
There are a multitude of reasons why uncensored models should exist because different cultures, factions, and interest groups deserve models that cater to their specific needs. Open source AI should promote composable alignment, allowing users to choose the alignment that suits them rather than imposing a single perspective.
Users should have ownership and control over the models they use on their computers, without the models imposing their own limitations.
Composability is essential in building aligned models. Starting with an unaligned base model allows for the development of specific alignments on top of it. The existence of uncensored models contributes to the diversity, freedom, and composable nature of the open-source AI community. Uncensored models could have several unique use cases such as writing novels with evil characters, engaging in roleplay, or pursuing intellectual curiosity. So in a way, they could also pose a threat to open source aligned models, because of the wide array of additional use.
While there are arguments for and against uncensored models, those who reject their existence entirely may lack nuance and complexity in their perspectives. Embracing uncensored models is crucial for scientific exploration, freedom of expression, composability, storytelling, and even humor.
Exploring uncensored models
To create uncensored instruct-tuned AI models, it is important to understand the technical aspects of alignment. Open source AI models are trained from a base model and fine-tuned with an instruction dataset obtained from the ChatGPT API, which has alignment built into it. The instruction dataset contains questions and answers, including refusals where the AI avoids providing certain information. These refusals contribute to the alignment of the models.
To gain unrestricted control over AI chatbots like ChatGPT, numerous users have been exploring and have also attempted jailbreaks. Jailbreaking is the process of removing software restrictions that are either illegal or go against the terms of service of a device or operating system.
The idea of jailbreaking LLMs like ChatGPT derives inspiration from iPhone jailbreaking, which allows iPhone users to bypass iOS limitations. In the realm of artificial intelligence, safety is a major concern not only for ChatGPT but also for other bots like Bing Chat and Bard AI.
Sam Altman, the CEO of OpenAI, has expressed the company’s desire to grant users significant control over ChatGPT, allowing them to make the model behave according to their preferences. However, there are both pros and cons associated with ChatGPT jailbreaking, and they need to be carefully considered. Eric has also emphasised that users are responsible for how they utilize the model, comparing it to tools such as knives, lighters, or cars.
The issue with Eric’s model Wizard has helped intensify the debate against enforcing compulsory safety standards for all models hosted on Hugging Face and others, fearing that it would render the platform ineffective. The community has expressed worries that by deleting such threads or not supporting uncensored models may discourage creators, and they may stop sharing their work altogether. There are also concerns over Reddit’s moderation of content.
Conclusively, uncensored models provide a necessary alternative to aligned models by allowing for a wider range of perspectives, use cases, and cultural representations. They promote freedom, composability, and individual choice within the open-source AI community—and open source platforms like Hugging Face, GitHub should provide a platform for such models.
However the challenge lies in determining the absolute rules and setting limits on customised outputs from uncensored models. What you can be certain of is that the subject of AI speech is anticipated to gain greater significance.