PHI-4-ABLITERATED

Introduction

Phi-4 Abliterated is a text generation model based on the Phi-4 architecture. It is designed to serve as a foundational model for various AI-powered features, with specific capabilities for reasoning and logic. The model is built to handle memory and compute-constrained environments, making it suitable for applications requiring low latency.

Architecture

Phi-4 is a 14 billion parameter dense decoder-only transformer model. This architecture allows it to perform complex text generation tasks efficiently.

Training

The training data for Phi-4 extends from its predecessor, Phi-3, and incorporates a variety of sources:

• Publicly available documents filtered for quality.
• Synthetic data designed to teach subjects like math and coding.
• Acquired academic books and Q&A datasets.
• High-quality chat format data for instruction adherence and preference alignment.

Safety measures include supervised fine-tuning and direct preference optimization using diverse datasets to enhance the model's robustness and adherence to instructions.

Guide: Running Locally

To run the Phi-4 Abliterated model locally, follow these steps:

1. Clone the Repository:
   Clone the model repository from the Hugging Face Model Hub or GitHub.

2. Install Dependencies:
   Ensure you have the necessary libraries installed, typically including transformers, torch, and other dependencies specified in the repository.

3. Load the Model:
   Use the Hugging Face transformers library to load the model and tokenizer.

4. Inference:
   Execute inference using the model for your specific text generation task.

Suggested Cloud GPUs

For optimal performance, consider using cloud-based GPUs such as those from AWS, Google Cloud, or Azure, which offer powerful hardware suitable for running large models like Phi-4 efficiently.

License

The Phi-4 Abliterated model is released under the GPL-3.0 license, allowing for open use and modification while ensuring that derivative works are also open-sourced under the same license.