Google

A prompt template for generating a new sample in the Self-Instruct process provides the LLM with existing samples to learn from. The template instructs the model to create a new sample, including an instruction, user input, and output. An example template is: "You are provided with a set of input-output samples, each composed of an instruction, a user input, and an output. Please generate a new sample based on these. Sample 1: $${\mathrm{instruction}1}$$ Input: $${\mathrm{user\text{-}input}1}$$ Output: $${\mathrm{output}1}$$ Sample 2: $${\mathrm{instruction}2}$$ Input: $${\mathrm{user\text{-}input}2}$$ Output: $${\mathrm{output}2}$$ New Sample: $${\mathrm{new\text{-}instruction}}$$".

Example of a Prompt Template for Sample Generation in Self-Instruct

A research team is using a large language model to automatically generate new training examples. Each example should consist of an 'instruction', a user 'input', and a corresponding 'output'. The team provides the model with several high-quality examples in this format, followed by a final instruction. However, they find that the model often just slightly rephrases one of the provided examples instead of creating a genuinely new one. 

Analyze the team's final instruction to the model, provided in the case study below. Explain why this instruction is likely causing the problem and propose a specific, revised instruction to resolve the issue.

Troubleshooting an Automated Data Generation Process

A developer is using a large language model to generate new training data. The goal is to create new samples, each consisting of an instruction, a user input, and an output. The developer provides the model with several high-quality examples and then uses the following prompt template to request a new one:

'Here are some examples. Make a new one.'

Identify two weaknesses of this prompt template and explain how you would improve it to generate better results.

Evaluating a Prompt Template for Data Generation

A research team is using an automated process to generate new training data. The process involves showing a language model a few high-quality examples, where each example consists of an 'instruction', a user 'input', and a corresponding 'output'. The goal is for the model to then create a completely new, well-formed example that follows the same three-part structure. Which of the following prompts, given to the model after the examples, would be most effective and precise for this task?

Learn Before

Related