Phi-3-Vision's integration aligns with Microsoft’s Responsible AI principles, incorporating safety post-training in different stages to enhance performance.
Safety training datasets include both text-only RAI datasets and in-house Multi-Modal RAI datasets covering various harm categories from public and internal benchmarks.
Phi-3-Vision's RAI performance was evaluated across internal and public benchmarks, showcasing notable improvements in safety and overall performance.
Results demonstrate that safety post-training can aid Phi-3-Vision in enhancing RAI performance across various categories.