Hello, i have integrated chatGPT in one of my website, when i type ‘hai’ or ‘hello’, rather than giving an acknowledging response it is giving some weird responses. But when i checked in playground it is giving proper results. I have set excatly the same parameters in my integration like temperature etc. But giving different results compared to the one in playground. Would like to know what i am doing wrong here. Thanks!
Do you mean you have integrated the OpenAI API into your website?
ChatGPT is currently a web-based application hosted by OpenAI (as it the Playground).
Yes, i just used simple python code provided in the openai website to test. But it is working fine, only issue as i mentioned is the responses, not getting actual responses what we would like to see, especially if we type just Hi,hello etc…
Yeah, I normally do not give OpenAI such open ended prompts because without more context, the API tends to hallucinate replies, such as:
irb(main):002:0> require_relative './lib/openai_completions'
=> true
irb(main):003:0> Completions.completion("Hello")
=> [",\n\nI am writing to you to inquire about the availability of a room in your hotel. I am planning to stay in your hotel for a few days and would like to know if you have any rooms available.\n\nI am looking for a room with a double bed, a private bathroom, and a balcony. I would also like to know if there are any additional amenities such as a mini-fridge, a microwave, or a coffee maker.\n\nI look forward to hearing from you soon.\n\nSincerely,\n[Your Name]"]
irb(main):004:0> Completions.completion("Hi")
=> [",\n\nI am trying to use the \"Create a new project\" feature in the JIRA Software Cloud.\n\nTo create a new project, you will need to log in to your JIRA Software Cloud account. Once you are logged in, you will see a \"Create\" button in the top right corner of the page. Click on this button and select \"Create a project\" from the drop-down menu.\n\nYou will then be taken to a page where you can select the type of project you want to create. You can choose from a variety of project templates, such as Scrum, Kanban, and Basic. Once you have selected the project type, you will be able to enter the project details, such as the project name, description, and key.\n\nOnce you have entered all the details, click the \"Create\" button at the bottom of the page. Your new project will then be created and you will be able to start working on it.\n\nI hope this helps.\n\nRegards,\n\nJohn"]
You can try fine-tuning, but fine-tuning probably will not solve this for you, as fine-tuning takes the results from the core engine / model and then tweaks the output during the final decoding process.
Thanks for the reply. But the funny thing is when you type hello in playground it gives the actual response we want like ‘Hello! How can I help you today?’ I would like to get the responses similar to that. So what should i do?
i have used excatly the same model and parmeters in playground, but still not getting results same as in playground. Is there anything to do with the api key i am using?
That is because (we think) the Playground appears to use a different underlying model than the API and ChatGPT also used a different model (or model version).
First, you need to do a survey of the models that each “system” you are working on is using.
- ChatGPT, the model, temperature, etc is?
- Playground, the exact model, temperature, etc is?
- You API completion, the exact model, temperature, etc is?
To get to the root cause of a situation, you must obtain situational knowledge.
HTH
1 Like
The same for us.
Just ask “what date is today” ? Answer is really creative…2020 and some time 2021…
On playground it give the right date:
Is the API just an old version engine?
It s no sense that paid account have less quality that free.
Maybe engine it s just for different purpuse or maybe we are doing something wrong.
Any indication or help for understand this situation can be appreacite
The same weird thing with “what date is today”, by the way in playgrouds it says 2020 and other inaccurate answers. But web-based chatGPT answers correct. Great to know it’s not only my issue
That is because so many retail ChatGPT customers (including the news media, etc) complained about the date being wrong that OpenAI simply pre-filters questions like that.
In other words, the date more-than-likely does not come from the model; but from the application, where there is a filter to answer this question before it goes to the model.
HTH
1 Like
any other pre-filtered phrases ?) It’s not good
I am sure there are many.
The GPT-3 model pre-training data has no way to know the system time (in real time), so that is an easy one to spot.
Also, all the moderating that goes on is also part of the pre-filtering. The prompts are rejected before they are forwarded to the model.
HTH