Level Up Your Teaching With Custom GPT

As Artificial Intelligence continues to evolve, educators are finding innovative ways to incorporate AI into their classrooms. One such tool is CustomGPT (Custom Generative Pre-trained Transformer) provided by OpenAI through ChatGPT. By creating CustomGPT models, teachers can customise AI to meet the unique needs of their students and classroom dynamics. In this article, we will explore what CustomGPT means, how to create it, and how to refine it based on personalised needs, with specific applications in academia.

To understand what CustomGPT is, it is important to briefly understand what ChatGPT is.

What is ChatGPT?

ChatGPT is a variant of the GPT (Generative Pre-trained Transformer) models, created by OpenAI, specifically made for conversational tasks. It can engage in dialogue, answer questions, and even generate content based on prompts. It is a pre-trained model that has a wide range of general knowledge and can perform various tasks. It started with text-based tasks only, but has since developed the ability to generate and analyse images and data as well.

What is CustomGPT?

A CustomGPT is a powerful tool that leverages the capabilities of general-purpose GPT models such as ChatGPT, and adapts them for specialised tasks, resulting in more accurate, relevant, and efficient outcomes customised to specific needs. This customisation allows innovative and practical applications across various fields, including education, healthcare, legal advice, customer service, and more.

How does a CustomGPT differ from general ChatGPT?

While using AI, following the standard structure of a prompt can yield better results, but the best results are achieved when the prompt is refined according to personal needs. In the same way, ChatGPT in general has a huge knowledge base, but it is not refined for personalised needs. CustomGPT allows for such personalisation. Now, let us explore how this ability to personalise a GPT can be leveraged in the educational setting.

How to Create a CustomGPT in OpenAI’s ChatGPT?

As of August 2024, creation of CustomGPT has been greatly streamlined.

• First open ChatGPT.com, and login using your credentials.
• At the top of the left pane, click on the “Explore GPT” button.
• On the right top, click the “+Create” button.

This will take you to the page where you can converse with a chatbot. As of August 2024, there are two options visible on top of the chatbot interface: “Create & Configure”. ‘Create’ feature lets us converse with a chatbot to create the CustomGPT. Here, all the information can be provided to the chatbot, including how we want the CustomGPT to behave, the task it needs to perform, and any other rules or instructions it should follow. It even allows you to create a display picture for the CustomGPT. Everything in the “Create” part works with the chatbot.

The other one is “Configure”. ‘Configure’ consists of all the settings of the CustomGPT that is being created. It has the settings for Name, Description, Instructions, Conversation Starters, Knowledge, Capabilities, and Actions. Let us now delve into each of these settings one by one.

What are the settings in CustomGPT?

‘Name’ and ‘Description’ are the obvious settings that need no explanation. ‘Instructions’ is the most important part in the creation of CustomGPT, which we will explore after looking at the other settings. ‘Conversation Starters’ are the example input prompts that a user who opens this CustomGPT should see initially. They are like guiding questions or example prompts which can be used with the CustomGPT. ‘Knowledge’ option allows us to upload the files with data or knowledge, which we want the GPT to refer to while performing the assigned tasks. ‘Capabilities’ includes 3 different capabilities that should be enabled for CustomGPT as of August 2024. There might be more in the future. But currently, three of them include Web Browsing, DALL·E Image Generation, and Code Interpreter & Data Analysis. We can choose to enable or disable these abilities in the CustomGPT we make. ‘Web Browsing’ allows the CustomGPT to search the web for information in real-time, accounting for updated answers and information in the response. ‘DALL·E Image Generation’ allows the CustomGPT to create images in the conversation using the latest image generation model available to it. And ‘Code Interpreter & Data Analysis’ allows the CustomGPT to run code, analyse data and files we upload, do math calculations, and other tasks related to data science. And the last one, ‘Actions’ lets GPT retrieve information or take actions outside of ChatGPT, which we will not delve into due to its technical requirements.

What is the ‘Instructions’ setting? And how to write them?

We will now look at the ‘Instructions’ setting, which is the key part of creating a CustomGPT. All the information we provide to the chatbot in the ‘Create’ feature gets refined and added into the ‘Instructions’. In ‘Configure’, it is possible to edit the ‘Instructions’ made from our conversations with the chatbot, or completely delete and rewrite them from scratch. ‘Instructions’ are the rule sets and guidelines that the CustomGPT has to follow while performing assigned tasks.

Writing instructions is similar to writing a prompt. We can follow the same structure that is used for writing a prompt i.e., giving a persona to GPT, giving the context, assigning a task, providing rule sets, defining output format, etc.

To understand this better using a real-life example, consider a teacher using generative AI to evaluate their students’ answer scripts and provide constructive feedback. Below is an example of “Instructions” that the teacher might use to make AI perform this task for them.

“Act as an experienced evaluator who has excelled in evaluating students’ descriptive answer scripts according to the rubrics and rules provided. You will be provided with the question, correct answer, rubric, and student’s answer. Your task is to evaluate the student’s answer based on the rubrics and correct answer provided, assign marks, and give constructive feedback to the student.”

These instructions can be expanded by adding different rules for different subjects and grades, and any other specification required based on the use case.

After creating this CustomGPT, we can provide the question, answer, rubrics, and student’s answer to get the evaluation and feedback. In this process, the input needs to be given in text format. This is easier if all the data is in digital form.

But, what if the student wrote the answers on a sheet of paper? Then just click a picture of that paper and upload it as input. GPT now has the ability to read and analyse pictures. This reduces the need to type or convert the answers into digital text.

Conclusion

Each Large Language Model (LLM) available in the market today has a huge knowledge base and set of abilities. Many MNCs and organisations are using these LLMs to build refined Small Language Models (SLMs) for their specific needs. Building those SLMs will require significant financial and processing resources. Those SLMs will be efficient and useful if the use case is substantial enough to justify the resources. However, for personal use cases such as the one discussed in this article, CustomGPT fills the void.

Co-authored by Sankhya and Jyothi, Prompt Analysts at Coschool. To learn more, visit www.coschool.ai