71–75 of 75 posts
Maybe a dumb question, but how do you train a model that has already been trained? What is that process like?
1. You lie, as many of these startups do. They are just adding a custom prompt to your input. 2. You send OpenAI your corpus and your money, and they merge it with their corpus and run their software.
I don't see in the code how it is referencing any model that was fine tuned on late night monologues? All is see in an API call to OpenAI's "text-ada-001" model. I'm interested in the fine tuning (I wasn't aware OpenAI was allowing GPT3 weights out to be tuned?) and how that is being factored into the code.
In terms of where it's incorporated in the code, the fine-tuned model id is stored in an environmental variable called `process.env.OPENAI_MODEL_ID` which then gets passed to the completion endpoint
Maybe a dumb question, but how do you train a model that has already been trained? What is that process like?
Process is very straightforward -- all the code (data cleaning, uploading, tracking can be found here: https://github.com/brendansudol/joke-gpt/blob/main/train.ipy...)
Earlier quoted context omitted.
1. You lie, as many of these startups do. They are just adding a custom prompt to your input. 2. You send OpenAI your corpus and your money, and they merge it with their corpus and run their software.
Would it maybe be possible to extract that prompt by doing prompt injection?
This is the essence of GPT. It perfectly captures style, but not substance. I naturally read the jokes in Jimmy Fallon's voice, before knowing it was fine-tuned. It really takes talent like his to pull off jokes like this every night.
I'd quite like to see someone with genuinely excellent one liner delivery go through an entire set of these almost-a-joke-jokes, just to see the audience react.