Dolphin3.0 Qwen2.5 0.5 B

Introduction

Dolphin 3.0 is part of a series of instruct-tuned models designed for general-purpose use, including coding, math, and function calling. It aims to give users control over system prompts, model versions, and data alignment, contrasting with proprietary models like ChatGPT and Claude.

Architecture

Dolphin 3.0-Qwen2.5-0.5B is part of the Dolphin series, leveraging the Qwen2.5 architecture. It is trained on diverse datasets and designed to operate as a local, customizable AI model.

Training

The model was trained using a variety of datasets, including OpenCoder-LLM, Microsoft's Orca datasets, and others. Sponsors provided high-performance hardware like L40s and H100 GPUs for training. A notable feature of this model is its flexibility regarding system prompts and alignment, allowing users to steer the model's behavior.

Guide: Running Locally

1. Setup Environment: Install dependencies using Python virtual environments or Docker.
2. Download Model: Obtain the Dolphin 3.0-Qwen2.5-0.5B model from Hugging Face.
3. Load Model: Use the Hugging Face Transformers library to load and initialize the model.
4. Run Inference: Implement a system prompt to customize responses and run inference using sample inputs.

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

License

The Dolphin 3.0-Qwen2.5-0.5B model is licensed under the Apache-2.0 License. For more details, refer to the license document.