Llama3

Llama3 is Meta's advanced large language model (LLM) designed to empower individuals, creators, researchers, and businesses of all sizes. This release includes model weights and starter code for pre-trained and instruction-tuned models, with parameter sizes ranging from 8B to 70B. The repository serves as a minimal example for loading Llama3 models and running inference, while more detailed examples are available in the Llama Cookbook.

Key Features

• Model Variants: Available in 8B to 70B parameter sizes, supporting sequence lengths up to 8192 tokens.
• Safety Focus: Includes PurpleLlama for safety risks and inference-time mitigations.
• Flexible Deployment: Supports both native and Hugging Face transformers formats.
• Agentic Applications: The llama-agentic-system enables the creation of sophisticated agentic applications.

Getting Started

To use Llama3, you need to:

1. Visit the Meta Llama website to request model access and accept the license.
2. Download the model weights using the provided download.sh script or via Hugging Face.
3. Set up a conda environment with PyTorch/CUDA and install dependencies.
4. Run inference using the provided example scripts.

Safety and Responsibility

Llama3 includes tools for responsible use, such as safety classifiers and a Responsible Use Guide. Users are encouraged to report any issues or unsafe outputs through designated channels.

Licensing

Llama3 is licensed for both research and commercial use, with an emphasis on ethical AI advancements. The accompanying Acceptable Use Policy outlines guidelines for responsible deployment.

For more detailed examples and integrations, refer to the Llama Cookbook.