DialoGPT-large starts repeating himself after a few exchanges

#1
by sbreit - opened

I don't know if anyone here has had the same results, but this model starts repeating the same response after a while, no matter what I query him.
Even the online chatbox example on the model page does that.
Is that a known problem from this model? Anyone knows if there is a workaround for this?

I found this to be an issue as well. I'm not sure what it could be as I am just a newbie messing around with models. It seems to happen after so many queries.

https://github.com/microsoft/DialoGPT/issues/45 - solved. But there's another issue: after about 10 questions it starts sending one symbol or something like "Mmmkay" or "so".

Trained on Reddit threads, this is not a defect, would say it's doing exactly what it's learned.

Same issue here.

Same here. Here's the convo:

Input a message to start chatting with microsoft/DialoGPT-large.
Hello, can we chat?
Sure, I'll PM you.
Thanks, are you made by microsoft?
No, I'm not.
who where you made by?
I made this
who made you? sorry for bad question
I made me
Who are you?
I am you.
I love it. Im you also, or someday we'll be togheter
I am you.
I am you.
I am you.
Fine
I am you.
Who is God?
I am you.
Ok
I am you.
Thank you for that. Do you have any pets?
I am you.
Can you say anything else?
I am you?
I don't think so, lol
I am you.
Your funny
I am you.
Stop lagging
I am you.
Bye =)
I am.

this is not a chatgpt by OpenAi right?

this is not a chatgpt by OpenAi right?

GPT-2 text generator - is OpenAI project
DialoGPT chat-bot - is Microsoft project

I mean, look at the difference between ChatGPT and the new Bing/Sydney

It will be great, if we combine ChatGPT + Sydney + Character.AI

That wouldn't be a bad idea at all, if we could capture the personality of Sydney somehow. Maybe by summarizing Reddit posts about her?? Lol

deleted
This comment has been hidden

Sign up or log in to comment