New Delhi
Microsoft has introduced its latest compact language model, Phi-2, designed by the Machine Learning Foundations team at Microsoft Research. Phi-2, a 2.7 billion-parameter language model, exhibits exceptional performance, matching or surpassing certain larger open-source models like Llama 2 with fewer than 13 billion parameters. In recent months, Microsoft has rolled out a series of small language models (SLMs) under the umbrella of Phi. The initial model, Phi-1, with 1.3 billion parameters, achieved state-of-the-art performance in Python coding, particularly on benchmarks like HumanEval and MBPP. Phi-2, the newest addition, demonstrates outstanding reasoning and language understanding capabilities. Microsoft emphasizes that Phi-2 is an ideal platform for researchers, offering opportunities for exploration in mechanistic interpretability, safety enhancements, and fine-tuning experiments across various tasks. The model is available in the Azure AI Studio model catalog, encouraging researchers to delve into language model research and development. Microsoft’s approach with the Phi models aims to explore whether performance on par with larger models, yet far from the frontier models, can be achieved at a smaller scale through strategic training choices such as data selection. As language models have grown to unprecedented sizes with hundreds of billions of parameters, Phi-2 represents a compelling effort to investigate whether similar capabilities can be achieved at a smaller scale. Microsoft has conducted extensive testing on commonly used prompts from the research community, aligning with expectations based on benchmark results. The tech giant continues to push the boundaries of language model capabilities and accessibility.
