Google

An alternative method for generating initial prompts with a Large Language Model (LLM) involves providing the model with examples of task inputs and their corresponding desired outputs. This approach relies on the LLM's ability to infer the task's underlying instruction from these concrete examples in order to formulate an effective prompt, rather than being given an abstract task description.

LLM-based Prompt Generation from Input-Output Examples

To have a Large Language Model infer and generate an instruction from examples, one can use a template that provides the input-output pairs explicitly. For instance:

```
You are provided with several input-output pairs for a task. Please write an instruction for performing this task.
Input: *input1*   Output: *output1*
Input: *input2*   Output: *output2*
...
```

This formatting allows the model to deduce the correct task instruction.

Example of a Prompt Template for Generating Instructions from Input-Output Pairs

A key capability of Large Language Models is their ability to infer the underlying instruction for a task when provided with a series of input-output examples. By analyzing the relationship between the inputs and their corresponding outputs, the model can deduce the transformation rule or goal of the task.

Instruction Inference from Input-Output Pairs

A developer wants an LLM to generate the instruction: 'Classify the sentiment of the text as Positive, Negative, or Neutral.' They provide the LLM with the following input-output examples to infer this instruction:

- Input: 'I love this new phone!' -> Output: 'Good'
- Input: 'The service was terrible.' -> Output: 'Bad'
- Input: 'The movie was okay, I guess.' -> Output: 'Fine'

However, the LLM generates a vague instruction like 'Analyze the text provided.' Analyze the developer's examples and explain the most likely reason for this outcome. Then, suggest a specific improvement to the output examples to help the LLM generate the desired instruction.

Diagnosing Issues in Instruction Inference

A developer is building a tool to process customer complaint emails. The goal is to have a language model automatically extract three specific pieces of information: (1) the core issue, (2) the product involved, and (3) the customer's requested action. The developer's initial attempts at writing a direct instruction have failed to produce consistent results.

Your task is to help the developer by applying the technique of generating an instruction from examples. Create three distinct input-output pairs that you would provide to a language model to help it infer the correct instruction for this task. For each pair, briefly justify why it is a good example.

Crafting Examples for Instruction Inference

A developer wants a language model to generate a clear instruction for the task of 'summarizing a long paragraph into a single, concise sentence.' To do this, they will provide the model with a set of input-output examples and ask it to infer the instruction. Which of the following sets of examples is most likely to result in the desired instruction?

Learn Before

Related