Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

TL;DR

This paper introduces a series of highly capable, small-scale language models, including phi-3-mini, that achieve performance comparable to larger models and can be deployed on phones, with extensions for multilingual, multimodal, and long-context capabilities.

Contribution

The paper presents new small and medium-sized language models with high performance, scalable training data, and enhanced capabilities for multilingual, multimodal, and long-context tasks.

Findings

phi-3-mini rivals GPT-3.5 on benchmarks

phi-3.5-MoE outperforms similar open-source models

phi-3.5-Vision handles multi-image and text prompts effectively

Abstract

We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. Our training dataset is a scaled-up version of the one used for phi-2, composed of heavily filtered publicly available web data and synthetic data. The model is also further aligned for robustness, safety, and chat format. We also provide parameter-scaling results with a 7B, 14B models trained for 4.8T tokens, called phi-3-small, phi-3-medium, both significantly more capable than phi-3-mini (e.g., respectively 75%, 78% on MMLU, and 8.7, 8.9 on MT-bench). To enhance multilingual, multimodal, and long-context capabilities, we introduce three models in the phi-3.5 series: phi-3.5-mini, phi-3.5-MoE, and phi-3.5-Vision. The phi-3.5-MoE, a 16 x 3.8B MoE model with 6.6 billion active parameters, achieves superior performance in language reasoning, math, and code tasks compared to other open-source models of similar scale, such as Llama 3.1 and the Mixtral series, and on par with Gemini-1.5-Flash and GPT-4o-mini. Meanwhile, phi-3.5-Vision, a 4.2 billion parameter model derived from phi-3.5-mini, excels in reasoning tasks and is adept at handling both single-image and text prompts, as well as multi-image and text prompts.