My attempt to finetune a safe companion chatbot (it partly worked)

Success: The chatbot avoids sex :+1:
Failure: The chatbot was eager to discuss how to build a grenade launcher :frowning: :boom:

Hi there, I’m working on a chatbot and I’m a newbie to openai and fine tuned models. With that said I saw your video and project and thought it was very interesting and helped me a lot in learning how to create my own dataset and first fine tune. After some attempts I noticed that the fine tune is not always strictly following the dataset. Do you have some tips to do so that the chat bot will answer questions about the dataset in a more tighter way? Thanks for your work, really inspiring :smiley:

Basically you just need more data. The larger the fine-tune dataset, the better the performance. Or more epochs but then you run the risk of over fitting

Planning to build a chatbot with a fine-tuned model, because I need factual information.

I’ll have a look at your video. Thank you!