An introduction to a basic prompt engineering strategy
Introduction
LLMs are fundamentally predictive models that only anticipate the following word in the sequence of words based on the provided context, despite their apparent superhuman skills.
As a result, their results are strongly reliant on the context that the users’ inputs give in addition to the enormous amounts of data that they are trained with.
Regular users of chatbots powered by LLM are aware of the significance of context. Whether they are created specifically for LLM or are publicly accessible services (like ChatGPT), chatbots will find it difficult to complete the more difficult tasks without enough context.
Here, we dig into one of the most basic ways used to help LLMs to answer prompts correctly: providing context within user prompts. This is often done out with 3 different methods: zero-shot prompting, one-shot prompting, and few-shot prompting.
Zero-Shot Prompting
If you’ve ever interacted with an LLM-powered chatbot, you’ve probably already used zero-shot prompting unknowingly, which involves relying only on an LLM’s pre-trained data to respond to a given user prompt. For instance, let’s assume that we used ChatGPT as a sentiment classifier. An input with zero-shot prompting would appear as follows:
Determine the sentiment of this sentence, user prompt.
“This basketball weighs a lot,” is the phrase.
The mood of the statement “This basketball has a lot of weight” is neutral, according to ChatGPT. Instead of conveying a good or negative message, it only gives the basketball’s weight in terms of facts.
Without any extra information, the LLM considers this sentence to be neutral when given the query.
While this approach typically produces appropriate answers, it could not be adequate if the question is particularly complicated or the standards for the answer are out of the ordinary.
