Google

In few-shot learning, the provided demonstrations create a pattern that explicitly maps example inputs to their corresponding outputs. This pattern serves as a template that the Large Language Model uses to infer the task and generate predictions for new, unseen inputs.

Input-Output Patterns in Few-Shot Learning

In the context of few-shot learning, the input-output patterns provided as demonstrations do not need to be complex; it is possible to guide a Large Language Model effectively using simpler pattern structures.

Use of Simpler Patterns in Few-Shot Learning

You want a language model to extract the main subject from a sentence. Write a prompt that includes at least three distinct examples demonstrating a clear input-output pattern for this task. Conclude your prompt with a new sentence for the model to process, formatted for completion.

Constructing an Effective Input-Output Pattern

A developer is using a large language model to extract company names from news headlines. They provide the following series of demonstrations in a prompt. Based on the established input-output pattern, what is the most likely output the model will generate for the final headline? Explain your reasoning.

Predicting Model Behavior from a Few-Shot Pattern

A user provides a language model with the following examples to teach it a new task:

`Input: Apple -> Output: A`
`Input: Banana -> Output: B`
`Input: Cherry -> Output: C`

When the user then provides the new input `Input: Grape`, the model responds with `Output: G`. The user was expecting the output to be the full word `Grape`.

Which of the following best explains why the model produced an unexpected result?

Learn Before

Related