Microsoft Phi-4 Model

Introduction

Phi-4 is a high-performance open model developed to enhance reasoning capabilities using a blend of synthetic data, public domain sources, and academic resources. The model's focus is on quality data and advanced reasoning, incorporating rigorous enhancement and alignment processes for precise instruction adherence and robust safety measures.

Architecture

Phi-4 is a dense decoder-only transformer model with 14 billion parameters.

Training

The training data for Phi-4 extends beyond its predecessor, Phi-3, including:

• Filtered public documents and high-quality educational content.
• Synthetic data designed for teaching math, coding, and reasoning.
• Academic books and Q&A datasets.
• High-quality chat data for instruct-following and truthfulness.

Multilingual data comprises about 8% of the training dataset, with a focus on English.

Guide: Running Locally

To run Phi-4 locally, follow these basic steps:

1. Clone the repository from the Hugging Face Model Hub.
2. Install dependencies using a package manager like pip.
3. Configure the environment to ensure compatibility with your hardware.
4. Run the model using a compatible framework like PyTorch or TensorFlow.

For optimal performance, consider using cloud GPUs provided by platforms such as AWS, Google Cloud, or Azure.

License

Phi-4 is released under the Microsoft Research License Agreement (msrla). Please refer to the LICENSE file for detailed terms and conditions.