This repository curates a collection of prompt examples for Meta's Llama models, targeting developers and researchers working with large language models. It aims to provide a comprehensive resource for effectively interacting with and leveraging Llama's capabilities through well-crafted prompts, including specific formats for Llama 2 and Llama 3.

How It Works

The repository details the specific prompt formatting required for Llama 2 and Llama 3 models. Llama 2 uses a [INST] <<SYS>> system_prompt <</SYS>> user_message [/INST] structure for single turns and appends model_answer for multi-turn conversations. Llama 3 employs a <|begin_of_text|><|start_header_id|>system<|end_header_id|> system_prompt<|eot_id|><|start_header_id|>user<|end_header_id|> user_msg<|eot_id|><|start_header_id|>assistant<|end_header_id|> model_answer<|eot_id|> format. These structured prompts are crucial for guiding the models' responses, enabling tasks like JSON output generation and role-playing.

Quick Start & Requirements

This is a collection of prompts, not a runnable application. To use these prompts, you will need access to Llama models and a framework that supports their execution (e.g., Hugging Face Transformers).

Highlighted Details

• Detailed prompt templates for Llama 2 and Llama 3, including system prompt examples.
• Guidance on achieving specific outputs like JSON.
• Examples for role-playing scenarios, code interpretation, and tweet sentiment analysis.
• Techniques for "jailbreaking" models and creating expert personas.

Maintenance & Community

This is a community-driven collection. Contributions are encouraged. Links to relevant Hugging Face blogs and GitHub repositories are provided for further context.

Licensing & Compatibility

The repository itself does not specify a license. However, the prompts are intended for use with Llama models, which have their own specific licensing terms from Meta. Users must adhere to the terms of service for the Llama models they employ.

Limitations & Caveats

This repository is a curated list of examples and does not provide executable code or a unified interface. Users must adapt and implement these prompts within their own Llama-based applications. The effectiveness of prompts can vary based on the specific Llama model version and fine-tuning.