System Information in LLM Inputs
System information, also referred to as a system prefix, is a sequence of tokens positioned at the beginning of an input to guide the behavior of a Large Language Model. For example, it can be used to instruct the model to act as a helpful assistant and to refrain from producing toxic content.
0
1
Tags
Ch.2 Generative Models - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
System Information in LLM Inputs
A developer is preparing a data sample to train a language model to act as a helpful, concise code reviewer. The complete input for the model is constructed by combining several pieces of text. Match each component name on the left with its corresponding text from the input sample on the right.
Diagnosing Fine-Tuning Input Flaws
A developer is creating a data sample to fine-tune a language model. The complete input for the model is shown below. Analyze the text and identify the part that constitutes the 'system information', which sets the overall behavior and constraints for the model.
Input Text: "You are a helpful travel agent. Your goal is to find the most affordable travel options in Europe. Do not suggest any destination outside of Europe. Based on the user's request, provide a 3-day itinerary. User request: I want to visit Italy for 3 days on a budget of $500."
Learn After
An e-commerce company wants to use a large language model as a customer service chatbot. The model must be polite, stay on the topic of the company's products, and decline to answer unrelated questions or invent information. Which of the following initial instructions would be most effective at guiding the model to meet all these requirements?
Troubleshooting an LLM-Powered Internal Knowledge Base
Crafting System Information for a Specialized Chatbot