Microsoft has recently unveiled a new lightweight AI language model called Phi-3 mini, which boasts 3.8 billion parameters trained on 3.3 trillion tokens. This latest edition offers similar performance to competitor products and was evaluated against academic benchmarks and internal testing, proving its ability to compete with models such as “Mixtral 8x7B and GPT-3.5”.
The Phi-3 mini model achieved a rating of 69 per cent in massive multitask language understanding (MMLU) and 8.38 in multi-turn (MT) benchmarks. Despite its capabilities, Microsoft acknowledged that Phi-3 mini is limited in size for certain tasks and lacks the capacity to store extensive factual knowledge.
However, Microsoft stressed that they developed Phi-3 mini in alignment with their responsible AI principles, with input from internal teams helping to curate additional datasets tailored to address specific insights. This approach resulted in significantly reduced harmful response rates, demonstrating Microsoft’s commitment to ethical AI development.
Microsoft revealed that they used a scaled-up dataset training approach with heavily filtered web and synthetic data to develop Phi-3 mini, building on the foundation set by Phi-2. The model is currently available through Microsoft’s Azure cloud platform, providing users with access to its advanced features and capabilities.
Overall, the introduction of Phi-3 mini marks an important milestone in Microsoft’s ongoing efforts to push the boundaries of AI technology while maintaining ethical standards.