Microsoft has made a significant move in the field of AI development by changing the license of its Phi-2 model from research to MIT. This major development will accelerate progress in building powerful yet safe AI technologies. The year 2024 is off to an incredible start with the announcement of Phi-2 going open-source.
About Phi-2
Phi-2 is a Transformer with 2.7 billion parameters. Using the same data sources as Phi-1.5, Microsoft trained it and augmented the training with a new data source containing various NLP synthetic texts and filtered websites. When assessed against benchmarks, Phi-2 showcased a nearly state-of-the-art performance among models with less than 13 billion parameters. The model has a context length of 2048 tokens.
To learn more about Microsoft Phi-2, check out our detailed blog highlighting its extraordinary capabilities. For technical details, visit HuggingFace Model Page.
Why Open Source Phi-2?
Unlike its predecessors like Claude, which were kept private, Phi-2 was released last year with a research-only license. Only academics and developers could experiment with the model for non-commercial purposes. However, its potential for real-world applications remained untapped.
The transition of Phi-2 to the MIT license is a significant milestone that brings several benefits to the open-source community. Firstly, it allows for commercial use of the model, enabling developers to integrate Phi-2 into their applications and products. This not only fosters innovation but also opens up new avenues for monetization.
The model and training code being publicly available will facilitate reproducible results. Challenges around reducing toxicity, handling nuanced instructions and mitigating bias can be tackled in an open, transparent and collaborative manner at scale.
Phi-2’s small size also enables on-device and edge applications on smartphones, IoT, robots, etc. Overall, this move signals Microsoft’s commitment to open research as the best approach for safely advancing conversational.
Conclusion
The transition of Phi-2 to an open-source model under the MIT license marks a significant milestone in the field of NLP. With its impressive performance, versatility, and accessibility, Phi-2 holds immense potential for developers, researchers, and the open-source community as a whole. This move allows them to use and modify Phi-2 for their projects freely. This will potentially lead to increased usage and further advancements in AI technology.