Best Practices for Prompting Llama 2 Chat Models
Intro
This article covers best practices for prompting Llama 2, a recently released language model from Google AI.
Prompt Template
Llama 2 uses a prompt template with the following structure: ``` INST Prompter Message INST Assistant Message ``` The "Prompter Message" is the text you provide to guide the model's response. The "Assistant Message" is the model's generated response.
One-to-Many Shot Learning Section
The prompt template also includes a "one-to-many shot learning" section, which looks like this: ```
INST Prompter Message
… [SEP]
INST Assistant Message
``` This section allows you to provide multiple examples of input and desired output, which helps the model learn your specific requirements.
Variable Mappings
The prompt template supports several variable mappings that you can use to customize the model's behavior. These mappings are: - `{prompt}`: The user's input. - `{assistant}`: The model's response. - `{shot}`: A single example from the one-to-many shot learning section.
Tips for Prompting
Here are some tips for prompting Llama 2 effectively: - **Be clear and concise.** Your prompts should be easy for the model to understand. - **Use natural language.** Don't use technical jargon or overly formal language. - **Provide context.** Give the model enough information to understand the context of your request. - **Use examples.** The one-to-many shot learning section can be very effective for teaching the model what you want. - **Experiment.** There is no one-size-fits-all approach to prompting. Experiment with different techniques to find what works best for you. By following these best practices, you can improve the quality of your interactions with Llama 2 and get the most out of this powerful language model.
Comments