Hi there! I just got access to the beta API and am excited to start tinkering with this technology! One thing I was wondering is if it is possible to train a Q&A chatbot to answer from a specific set of articles (200+).
Currently, there are details for training a customer support chatbot here but that seems to take in data regarding past chat history from users. For someone who does not have thousands of lines of past conversations, this doesn’t seem very practical so I’m hoping to provide GPT-3 more details specific to a certain area (e.g. economics) in the form of plain text/articles.
Guidance on this is greatly appreciated! Just getting started so I might have missed this functionality if it already exists.
In this case, @m-a.schenk I believe the student would be the only player. They would ask questions to the chatbot regarding questions in a specific domain. For example, the economics chatbot will provide answers to questions students might ask such as “what are the limitations of the X model vs the Y model”
Got it, makes sense! Thanks for the clarification.
What is your take on this? Is it better to fine tune using the Q&A format they have provided or would it be better to fine tune using a large set of articles and other text relating to the subject matter? Is there a resource I can refer to that discusses fine tuning using just large amounts of text?
I’m using raw material (like KB articles) and GPT-3 to synthesize data. Also there’s a fantastic dataset from Stack Exchange. I searched Kaggle Datasets to find tons of data. The short version is that you find lots of data that is “close enough” to bootstrap your project and then you use GPT-3 to synthesize the rest of the fine-tuning dataset.
Hi @ayangupta I just published some code and a tutorial for doing what you’re asking about. I used the answers endpoint and a documents file as the source for the answers. It’s a complete app that you can get up and running in a few clicks. Also, all of the code is there for you to look at / use / modify / whatever. Here is the tutorial video. I hope this is helpful.
