Is anyone facign a problem with custom gpts where the gpts are repeating the prompts or questions

Is anyone facign a problem with custom gpts where the gpts are repeating the prompts or questions? I have created a certain professional servies GPT which asks multiple questions to the users. For the last 12+ hours, GPT is repeating the questions and not following the right sequence

2 Likes

@ashish1109 Custom GPTs may be broken :frowning:

Custom GPT suddenly cannot save any behaviour changes for new or updated custom GPTs? - GPT builders - OpenAI Developer Forum

1 Like

Might be related to this:

Yep, having the same problem, plus the GPT is simply not performing the tasks in the config.

Looks like they are pretty broken today!

1 Like

This happens because of ambiguities in the natural language with which we give instructions. My advice is to use structured JSON files where the bot is told what to do and what not to do. JSON is more direct with the instructions that the bot should execute, because it is the most effective way to communicate instructions.

3 Likes

@iris3dgames @Kepulikeppi @Bizualize @ashish1109

Yes, I am getting this, so my CustomGPTs are pretty useless.

It seems they are broken further - Custom GPT suddenly cannot save any behaviour changes for new or updated custom GPTs? - GPT builders - OpenAI Developer Forum

Yes. I built a custom gpt in the playground and it worked perfectly, start to finish. Then went to the homepage environment and ran it…it totally crashed.

Next day went back to run the model in the playground where it worked perfectly yesterday and it totally crashed. Despite my not changing anything.

I believe that actual developers must be making the ones that work? Or are they being produced in the API? Not sure how the gpt store has 3 million apps already.

1 Like

And yes, it repeats prompts. Sometimes gets caught in a loop. I see one person suggesting we write in JSON. I’ll try, but we’re not supposed to have to. I only jumped in here because I’m not a coder and don’t want to be.

1 Like

JSON is the most structured way to tell the bot what to do or what not to do, it’s not that complicated, you upload the json file to the builider gpt and in natural language you tell the bot. Read the instructions of Start.json.

The point is that the structure of a json file helps that there are no typos in what the bot should do, they are clear instructions of what it should do or what it should not do.

For example, you can tell the bot, I want you to be an expert in mathematics, it is a perfect instruction, but if you ask the bot to give you a food recipe, it will do it, so you must be clear on how it will do things, train it step by step and this way it will work as it really wants.

And don’t worry, if you are not a coder, the gpt builder has what you need to create something interesting.

Yes, similar problem, read my account below:

HELP PLEASE … I am using the ChatGPT 4 Builder to simply create a Career Improvement Class. The model simply has to present, score and interpret 8 short questionnaires that make up the class. I have been struggling for a month. Two weeks ago, it successfully presented, scored and interpreted all 8 questionnaires properly. But since then, it presents the questions in the wrong order, or in the correct order with the wrong number attached to the question, or sometimes the same question twice. Mind you, i have not changed the files i upload and I have not changed the instructions to the gpt model i pasted into the Configure window. All conditions on my end are the same. I write everything in plain english text, because I’m supposed to be able to, right? Can someone help me determine what is going on here? Thanks. Jim

It seems this stuff is happening everywhere and to so many. I’m constantly rewriting my prompts. Since i posted this 10 hours ago, I was able to get my gpt model to go through the entire course, but it threw in errors at times. I made a record of the entire day on paper and have to analyze it. I wish there were some ‘pros’ I could hire to look my stuff over…

maybe if I wrote in python? What are you writing yours in ?

I built a Crypto Jurisdiction Finder on CustomGPT where I assessed 80+ countries on 30+ parameters. Check my GPTs on the marketplace. Everything was done via NLP with Retrieval. I ensured the data in the document that was uploaded on Retrieval was checked 50 times for accuracy, and tried instructions 10+ times. The CustomGPT is now working perfectly. Its now being used by Crypto & Web3 lawyers across the world. Looks like CustomGPT is fixed to a large extent. Keeping my fingers crossed.

Hey @ashish1109 ,
I develop a lot of custom GPT’s and from my Past usage couple of my GPT’s have went underrated.

What I did is I just finetuned the System Prompt again and it responded again in a Proper manner.

What do you mean with “I just finetuned the System Prompt again”? Is this normal training?

“Fine-tuned” in this case is referring to the act of iteratively adjusting the prompt wording to achieve the desired result, rather than model fine-tuning with training data.

2 Likes

yes @jlages94 you got the answer from @Foxabilo
I have fintuned my model just by adjusting prompt again and again.

1 Like