Deepy 3000 - no response after a few conversation turns

Hello, I am working on my own version of Deepy 3000 with DialoGPT implemented as a skill. I am currently running the model with one skill only. I noticed that after a while the model doesn’t produce dialogue replies anymore. Is it because I am hitting a hardcoded limit? Thank you for your help!

Here are the experiments I made:

  • Empty reply after 5 conversation turns, 138 tokens (measured using the pretrained DialoGPT tokenizer from HuggingFace).
  • Empty reply after 6 conversation turns, 131 tokens
  • Empty reply after 8 conversation turns, 223 tokens.

Hi, and welcome back!

Well, we don’t have any limitations built into the DP Agent that powers Deepy 3000. However, your DialoGPT model might have. Maybe not limitations per ce but maybe more context or others things needed for it to work properly. I would advise you to work with the creators of the model…

Thanks for your reply! The reason why I wrote here is that if I run DialoGPT alone (using model.generate, see the HuggingFace card) things work as expected, while if I implement it as a skill in DeepPavlov, after a while I get empty replies. Anyway, I’ll check things better on the DialoGPT side. Thanks again!

For posterity: I didn’t find a way to avoid getting empty replies when using the DialoGPT model from Microsoft, but when I used the model which I fine-tuned by myself everything works fine! I guess they didn’t preprocess the input data very carefully. Problem solved :heart_eyes: