Phi-3Translation site

Phi-3 is a large language model released by Microsoft with distinctive features and excellent performance.

1. Version and Parameter Scale::
   • There are three versions of Phi-3, phi-3-mini, phi-3-small, and phi-3-medium, corresponding to parameter sizes of 3.8B (3.8B), 7B, and 14B, respectively.
   • Among them, the phi-3-mini version is pre-trained on 3.3T tokens (tokens) of data, while the last two versions are trained with 4.8T tokens of data.
2. Technical characteristics::
   • Support for context windows of different lengths: Phi-3-mini is the first model of its kind to support up to 128K token context windows.
   • Command AdjustmentPhi-3 has been adapted to follow different types of commands, making the model ready to use.
   • Cross-platform operation: Phi-3 supports cross-platform ONNX Runtime and runs on GPUs, CPUs and even mobile hardware.
   • NVIDIA Optimization: Phi-3 is offered as an NVIDIA NIM microservice with a standard API interface that can be deployed anywhere and is optimized for NVIDIA GPUs.
   • Training data modulation: One of the core technologies of Phi-3 is to highlight the common-sense reasoning and logical inference properties of the training data by modulating the data and utilizing data filtering techniques and data synthesis techniques based on the large model, so that the small model can achieve NLU (Natural Language Understanding) and NLI (Natural Language Inference) capabilities comparable to those of the large model.
3. performance::
   • Phi-3-mini, despite being optimized for deployment on cell phones, has performance comparable to models such as Mixtral 8x7B and GPT-3.5.
   • In the MMLU (Massive Multi-task Language Understanding) benchmark, Phi-3-mini achieved an accuracy of 69%; in the MT-bench benchmark, it achieved a score of 8.38.
4. Safety and Responsibility::
   • Phi-3's training process includes post-safety training such as reinforcement learning human feedback, automated testing, and evaluation of dozens of hazard classes.
   • The model is compliant with the Microsoft Responsible AI Standard, following the principles of accountability, transparency, fairness, reliability and security, privacy and safety, and inclusiveness.
5. application scenario::
   • The Phi-3 family of models (including Phi-3-vision) plays an important role in real-world application scenarios due to its multimodal characteristics, such as for enhancing the multimodal conversational interface of Farmer.Chat and for helping medical software company Epic summarize complex medical histories more efficiently.

Phi-3 is a high-performance large-scale language model introduced by Microsoft, which not only has excellent technical features and performance, but also has a high degree of security and accountability, while being suitable for a variety of application scenarios.