Microsoft pulls WizardLM-2 AI model due to missed toxicity testing

Microsoft’s highly anticipated release of the WizardLM-2 AI model hit a snag as the tech giant swiftly removed it from public access mere hours after its launch. The successor to the renowned WizardLM AI model was retracted from both GitHub and HuggingFace, leaving behind 404 error links and a void in Microsoft’s official collections.

In a statement issued on X, the developers expressed regret, acknowledging their oversight in omitting toxicity testing from the model’s release process. Assuring the community of swift action, they pledged to conduct the necessary tests promptly before re-releasing the model.

Questions arose regarding the nature of the now-deleted models, with WizardLM-2’s developers clarifying that the omission stemmed from forgetfulness rather than any deliberate attempt to bypass scrutiny. The revelation that the research group is based in Beijing, China, adds a layer of context to the incident.

Despite the removal, remnants of the model persist online, with some users claiming to have downloaded and reuploaded the files. Concerns regarding potential risks associated with these unofficial versions linger, prompting scrutiny of Microsoft’s forthcoming toxicity tests.

Hugging Face and its CEO, Clément Delangue, expressed disappointment over the removal, highlighting WizardLM’s open-source models’ significant impact on their platform. They are actively seeking a resolution with Microsoft to address community needs.

As the saga unfolds, attention turns to Microsoft’s commitment to responsible AI practices. While the company declined to comment directly, the release of its updated Responsible AI Standard General Requirements underscores its dedication to ethical AI development, emphasizing the need to mitigate biases and disparities in AI outputs.