I am currently writing my master thesis and am experimenting with GPT-3. I got the task to write an experimental medical chatbot (I know about the rules, its just for research) and evaluate its performances based on the data given.
The problem, which i am currently facing, is how to fine tune GPT-3 with my datasets. I read the fine tuning article about 20 times and figured out how to promt engeneere, but what i am still lacking is the in and output of the chatbot. I want to demonstrate the chatbot using some sort of small GUI on my pc, and not within the playground.
Does anyone have a solution to this or can explain to me how i can get a constant chat on a python terminal on my pc?
Using “An Efficient DP-SGD Mechanism for Large Scale NLP Models” or some thing to interact with the model is very important. “Prompt modelling” is not sutainable or stable.
On top of all of this also look into choosing a cognitive architecture, SPAUN, ACT-R, NLCA, SOAR. You need a cognitive architecture because on itself the Transformer decoder is not going to magically behave very well.
My preference is do it in Jupyter but the output is your preference. Jupyter has a learning curve but it is a lot easier for production.