Increasing the Quality of Your GPT Instructions

Increasing the Quality of Your GPT Instructions

We are constantly on the lookout for things that can increase our quality even more. Recently, we saw a YouTube video made by WesGPT, where he allegedly is able to reverse engineer OpenAI's own system message to ChatGPT. Whether you think it spat out the actual system message or if it's mumbo jumbo, we decided to do some testing to see if we could increase the quality of our own system messages using this formatting.

So what's new?

What's the biggest difference between how we used to format the system messages and the new way? There are two main components: how the list is formatted (with // and it's numbered instead of bulleted), and using CAPS LOCK for anything that is extra important, such as "DO NOT" or "ALWAYS".

Last night, I created two chatbots that were trained on the exact same data, but one has the old system message and one has the new one. Both are pretty short and concise, but they have the same rules, just formatted differently. See images for comparisons.

Here is the old system message formatting:

Screenshot of the old system message

Here is the new system message formatting:

Screenshot of the new system message

The quality difference

While the quality depends a lot upon the data your model is trained on, we were pretty amazed by the results. Some questions will return almost identical replies (for example "show me images of your founders"), but anything that needs your model to do creative writing (while still adhering to facts) looks a lot better to us with the new system message. We'll go through a couple of examples, but keep in mind that while both of these models are trained on the exact same data, the training data is a bit limited.

Here is how it replied to the question "What is AINIRO?" in the model using the old formatting. While this is a pretty good reply, this isn't really a perfect reply.

Screenshot of the old system message reply

Next is how the model using the new formatting replied to the exact same question. Instead of using a random image, it's using our logo, and we found that the reply itself is also higher quality.

Screenshot of the new system message reply

Next is how the model using the old formatting replied to the question "What is this company?".

Screenshot of the second old system message reply

And lastly, how the model using the new formatting replied to the same question. It's using a more relevant image, it's answering the question of what we can do for the user, and generally it's a reply we're more happy with.

Screenshot of the second new system message reply

Changing the system message should only take you a couple of minutes, as you can use the same rules you already have, just reformatting them slightly. One thing I did notice was that in some cases I needed two rules for the same thing, for example "display images and hyperlinks as MARKDOWN" didn't prevent AI hallucinations. When I added another rule mentioning "DO NOT make up images and hyperlinks not found in the CONTEXT", AI hallucinations were eliminated again. So if you want to change your system message for higher quality, you do need to do some testing, and you may need to add 1-3 extra rules, but the results are definitely worth it in our opinion.

Conclusion

You're of course free to make up your own mind on whether you agree with us that the new formatting is better, but we definitely prefer the way it replies using the new formatting, and so far, all clients we've tested it with have agreed with us. It's using even more relevant images, it's not just answering the question directly but also adding arguments for what we can do for the user, and it's wording and formatting the replies better.

What else can this be used for? This can be used for all of our AI products, and you can use it for your own private GPT's on OpenAI's platform, whether it's connected to a Cloudlet or not.

Aria Natali Aurora

Aria Natali Aurora Nygård I am the COO and Co-Founder of AINIRO.IO AS together with Tage. I write about Machine Learning, AI, and how to help organizations adopt said technologies. You can follow me on LinkedIn if you want to read more of what I write.

Published 7. Feb 2024