Generating and editing prompt

A good prompt provides the context to the model to enable the best judgement from it. It puts all the context parts into a single place. Task description, class names and descriptions, dataset items, and other data can be used to build a good prompt.

Toloka LLM splits the provided dataset into two parts:

  • Measure quality — Items labeled by a model for quality measurement.
  • Prompt examples — Items that will be used as few-shots to show a model what you want it to do.

Providing few-shot examples significantly improves the result quality. At the same time, to get a representative measurement Toloka LLM doesn't include examples into a dataset used to measure quality on.

Provide few-shots

The model selects some texts to use as few-shot examples.

To edit the selected examples or provide more few-shot examples, click the Edit few-shots button. The Review few-shots for the prompt popup will open.

Mark the texts to use them as few-shots. The Distribution of few-shot examples area shows the ratio between the classes in few-shots. If you see that some of the classes have a very small percentage, it might be a sign that you need to select more text examples for this class to use as few-shots for better model results.

The selected few-shot texts will be used by the model in the prompt.

Customizing the prompt

You can use variables to customize your prompt. Variables allow you to include data from your dataset file into the prompt. They will update together with the dataset file and will not have to change the prompt manually.

The variables include:

  • {task} — The description of the task you provided.
  • {class.name}, {class.description} — The details of the classes you created or uploaded in the dataset file.
  • {example.text} — The text from the uploaded dataset file which you selected as a few-shot.

Additionally, you can include some tags into your prompt (like, <text></text> or <example1></example1>) to further structure the prompt for the model.

Available models

You can use the following models to generate prompts:

  • GPT 3.5
  • GPT 4

To change the model used, click the Prompt powered by drop-down list and choose the model for your prompt.

When ready, click Generate prompt Generate prompt.

Next steps

Last updated: September 5, 2023

Introduction
Toloka LLMRegistering and signing inAdding OpenAI API keyTerminology
Working with Toloka LLM
Workflow overview
Working with datasets
Problem definition
Iterating on quality
Creating endpoint