Join our daily and weekly newsletters for the latest updates and the exclusive content on AI coverage. Learn more
In the race for the deployment of corporate AI, an obstacle systematically blocks the path: hallucinations. These responses made of AI systems have provoked everything, legal sanctions for lawyers with companies forced to honor fictitious policies.
Organizations have tried different approaches to Resolve the hallucination challenge, including fine adjustment with better data, increased generation (RAG) of recovery and railings. Open source development company Oumi Now offers a new approach, although with a somewhat “cheesy” name.
THE The name of the company is an acronym of Open Universal Machine Intelligence (OUMI). It is managed for ex-Apple and Google Engineers on a mission to build an unconditional open source AI platform.
On April 2, the company published Halloumi, an open source complaint verification model designed to solve the precision problem thanks to a new approach to hallucination detection. Halloumi is, of course, a type of hard cheese, but that has nothing to do with the name of the model. The name is a combination of hallucination and Oumi, although the time of the release near the day of the April fools could suspect that the release was a joke – but that is anything but a joke; It is a solution to a very real problem.
“Hallucinations are frequently cited as one of the most critical challenges of the deployment of generative models,” said Manos Koukoumidis, CEO of Oumi, in Venturebeat. “This ultimately comes down to a question of confidence – generative models are trained to produce outings that are probablely likely, but not necessarily true.”
How Halloumi works to resolve the Hallucinations of the AI of business
Halloumi analyzes the content generated by AI on a sentence by sensation. The system accepts both a source document and an AI response, then determines whether the source equipment supports each assertion in the response.
“What Halloumi is doing is analyzing each sentence independently,” said Koukoumidis. “For each sentence he analyzes, he tells you the specific sentences in the input document that you must check, so you don’t need to read the entire document to check if [large language model] He said is precise or not. “”
The model provides three key outputs for each sentence analyzed:
- A trust score indicating the probability of hallucination.
- Specific quotes linking claims to supporting evidence.
- An explanation readable by man detailing why the complaint is supported or not sustained.
“We trained him to be very nuanced,” said Koukoumidis. “Even for our linguists, when the model signals something like a hallucination, we initially think that it looks correct. Then, when you look at the justification, Halloumi emphasizes exactly the nuanced reason why it is a hallucination – why the model made a kind of insurance, or why it is inaccurate in a very nuanced way.”
Integrate Halloumi into business AI workflows
There are several ways that Halloumi can be used and integrated into the AI company today.
An option is to try the model using a somewhat manual process, although the online line Demonstration interface.
An API -centered approach will be more optimal for the production and workflows of corporate AI. Manos explained that the model is fully open-source and can be connected to existing workflows, run locally or in the cloud and used with any LLM.
The process consists in nourishing the original context and the LLM response to Halloumi, which then checks the output. Companies can integrate Halloumi to add a layer of verification to their AI systems, helping to detect and prevent hallucinations in the content generated by AI.
Oumi has published two versions: the generative 8B model which provides a detailed analysis and a classifier model which provides only a score but with greater calculation efficiency.
Halloumi vs rag vs railings for the protection of Hallucinations of AI of business
What distinguishes Halloumi from other grounding approaches is the way it completes rather than replacing existing techniques such as the cloth (increased recovery generation) while offering a more detailed analysis than typical guards.
“The entry document that you feed the LLM could be a cloth,” said Koukoumidis. “In some other cases, it is not precisely Rag, because people say:” I do not recover anything. I already have the document that is close to my heart. I tell you, it is the document that is close to my heart. Summary for me. “Thus, Halloumi can apply to the cloth but not only to the cloth scenarios.”
This distinction is important because if RAG aims to improve generation by providing a relevant context, Halloumi checks production after generation, whatever the way this context was obtained.
Compared to railings, Halloumi provides more than binary verification. His analysis at the level of the sentence with scores and explanations of confidence gives users a detailed understanding of the place and the way in which hallucinations occur.
Halloumi incorporates a specialized form of reasoning in his approach.
“There was certainly a variant of reasoning that we made to synthesize the data,” said Koukoumidis. “We have guided the model to reason step by step or claim by sub-statement, to reflect on how it should classify a greater complaint or a greater sentence to make the prediction.”
The model can also detect not only accidental hallucinations, but intentional disinformation. In a demonstration, Koukoumidis showed how Halloumi was identified when the Deepseek model ignored Wikipedia content and rather generated propaganda content on the COVVI-19 response from China.
What it means for the adoption of corporate AI
For companies that seek to open the way to the adoption of AI, Halloumi offers a potentially crucial tool to securely deploy generative AI systems in production environments.
“I really hope it unlocks many scenarios,” said Koukoumidis. “Many companies cannot trust their models because existing implementations were not very ergonomic or effective. I hope Halloumi will allow them to trust their LLM because they now have something to breathe the confidence they need. ”
For companies on a slower AI adoption curve, the open source nature of Halloumi means that they can experience technology now while Oumi offers commercial support options if necessary.
“If companies want to better personalize Halloumi in their field, or have a specific commercial way, they should use it, we are always very happy to help them develop the solution,” added Koukoumidis.
While AI systems continue to move forward, tools like Halloumi can become standard components of corporate AI batteries – essential infrastructure to separate the facts of fiction AI.