Hello! I’m working on making a discord chat bot using the OpenAI API. I want the bot to function similarly to chat GPT in the sense that it can take a random input from a user and return something coherent, however I want to introduce a very large dataset to serve as its “memory” which far exceeds the token limit. At first I thought that this would be a job for the fine-tuning tools but I then learned that text-davinci-003 isn’t available for fine tuning. Does this mean I’m out of luck, or is there a creative solution that could be used to train the standard davinci to work as an all-around chatbot with a specific memory?
To be specific, I documented everything I did every day for a year, and I want to be able to talk to the bot about that year, and for it to write in the style that I wrote all the entries in. I attempted to create a fine-tune where every prompt started with “What did you do on [date]” and the completion being the text of that day, but based on the gibberish responses I’m receiving, it’s clear to me that I’m missing the point of the fine tuning, and I’m not surprised because I’m very much a beginner! I would love some advice on what route to take or if what I’m trying to do is even possible with the current tools available and limited experience. Thank you!
Hey i think you are a bit confused on the terms, im doing something similar for my company as a proof of concept with the tech support documentation.
What you need to use are embeddings , you need to “upload” the data you want to search/question as vectors and then obtain the similarities , to get a response as close to your answer as possible.
Here’s a youtube link with an excellent tutorial:
@piers.batchelor Its seems your use case may not be a chat bot Would that be right?
By reading your description it sound like you want to take a document and do some sort of analysis on its content. Is this correct, or are you are trying to ask questions about parts of the document (They require different solutions)
a) Chatbot trained on large body of text
b) Analyse content of a large body of text
c) ask questions about a large body of text (similar to a - but no back and forth discussion)