Google

A direct, imperative command is a technique used in prompting to strictly limit a Large Language Model's output to a predefined set of options. For instance, the instruction 'Just answer: positive, negative, or neutral' forces the model to choose its response from only those three words, ensuring the output is structured and easily parsable for classification tasks.

Constraining LLM Output with a Direct Command

A prompt for polarity classification can be structured by placing the instruction after the input text and question. This format presents the content for analysis first, followed by a direct question and a final command that constrains the output to a specific set of labels. For example:

'The service at the restaurant was slower than expected, which was a bit frustrating.

What is the polarity of the text?

Just answer: positive, negative, or neutral.'

Example of a Prompt with a Post-Input Instruction for Polarity Classification

A data scientist is using a large language model to classify thousands of user product reviews. For their data processing pipeline to work correctly, the model's output for each review must be *only* one of the following three words: 'satisfied', 'dissatisfied', or 'neutral'. Write a single, imperative sentence that you would add to the end of their prompt to ensure the model provides the output in this exact format.

Crafting a Constraining Command for an LLM

A data scientist is using a Large Language Model to automatically label thousands of customer reviews for a sentiment analysis project. Their goal is to create a dataset where each review is tagged with one of three exact labels: 'Positive', 'Negative', or 'Neutral'. However, their automated script keeps failing. They discover the model is producing a wide variety of responses, such as 'The sentiment of this text is clearly positive.', 'I'd say this is a negative review.', or simply 'Neutral'. Based on this information, analyze the likely cause of the script's failure and describe the specific modification you would make to the prompt to ensure the model produces consistent, machine-readable output.

Diagnosing a Failing Data Labeling Pipeline

A developer is building an automated system to sort customer support tickets. The system requires a language model to categorize each ticket's sentiment so it can be routed correctly. For the system to work, the model's output must be *exactly* one of the following words: 'Positive', 'Negative', or 'Neutral'. Which prompt most effectively uses a direct, imperative command to achieve this specific, constrained output?

Learn Before

Related