According to Hoshio, last week, the big company Microsoft, which is known as one of the most influential players in the field of artificial intelligence, released the latest version of its collection of small language models (SLM) named Phi-2 in its catalog of models at Azure AI Studio, made available to the public. This small language model is currently only usable for research purposes.
SLM or Small Language Model is a type of language model that is developed and trained only on a specific domain, and this makes it an ideal option for academic use and achieving scientific and research goals. This model is also more cost-effective than large language models due to its smaller size and is designed and produced for tasks that do not require huge computing power.
One of the significant strengths of Microsoft's new language model, Phi-2, is the focus on textbook-based content quality for education. Microsoft explained that “Phi-2, with its small size, is an ideal field for research by researchers, including exploring mechanical interpretability, improving safety, and conducting detailed tests on a variety of projects.” The company went on to state that, “Our educational data mix consists of synthetic datasets specifically created for teaching reasoning and general knowledge including science, daily activities and theory of mind, among others. We augment our educational portfolio with carefully selected data from the web, which is actually filtered based on educational value and content quality.”
Although Phi-2 is a small language model, it has 2.7 billion parameters and is a big improvement over the company's previous model, Phi-1.5, which only had 1.3 billion parameters. In addition, considering Microsoft's special approach to scaling, it may be possible to accept the company's claim that Phi-2 performs better than models that are 25 times larger. Microsoft's benchmark tests show that Phi-2 with its 2.7 billion parameters outperforms its two main competitors, SLM Mistrial with 7 billion parameters and Llama-2 with 70-7 billion parameters, in joint reasoning, language comprehension and solving mathematical questions. he does. The company also claims that its new model, despite its smaller size, can compete with the recently announced Google Gemini Nano 2. This model, like other small language models, is trained on specific data and is not enhanced by human feedback, which is why it has seen a reduction in bias compared to Llama-2.
Last week, Microsoft's new small language model called Phi-2 was released. This small language model has 2.7 billion parameters and is a significant improvement over the previous model of the same company with 1.3 billion parameters. Microsoft claims that the Phi-2 will provide better performance than its competitors, and even better or equal performance compared to Google's new model. This model is now available in Microsoft's catalog of models in Azure AI Studio.
What do you think? Will the Phi-2 be more powerful or the Google Gemini Nano 2?
Latest Passing over countries : Spain | Dominica | United Arab Emirates
RCO NEWS
