I'm trying to teach DialoGPT my style of conversation, but it's failing

Intermediate
1

I used ChatGPT and DeepSeek to create a trainer that will teach DIaloGPT-large my style of conversation. I was fine-tuning it, changing epoch, and slowing down learning. I have 7k of my own messages in my own style. I also checked my training dataset to be in the correct format.

But my model gives me stupid non-sense replies. They should ad least make some sense, since DialoGPT knows how to converse but it needs to converse in my style. What I’m doing wrong?

Here is my code python-ai-sexting/train.py at main · trbsi/python-ai-sexting · GitHub
My niche is specific and replies should be also

1 Like
2

For newer LLM models released in the past two years or so, simply using apply_chat_template (including implicit application) often suffices for the template to be incorporated without issues and fine-tuned in Trainer. However, since DialoGPT likely represents a model from a time when this aspect was still underdeveloped, explicitly handling it may be necessary.

1 Like
3

Thank you for this great info, I will check it out

1 Like