EDIT: Ao far @Alan said that he was successful with using prompts with 3 adjectives (so nearly zero tokens as input). We would like to hear what others have done with fine tuning and whether zero-prompt is possible. Maybe @boris know wether it is stiil good practice and can point to the case study he was talking about
Yes, you have to finetune your own model to do that. So in our case we came up with 100s of examples where 3 adjectives were used to create the paragraphs of text that we wanted. We hand-wrote all of those, then trained the finetuned model on it. It worked extremely well, so we were able to go from large prompts to produce that text, to just three adjectives. It can save a lot of money on text generation if you have something specific in mind.
How precise were the completions after fine tuning?
My use case is this:
I have a text like the one below:
San Diego’s Friendship Park, located on the U.S.-Mexico border, is one of them, as it has served for decades as a gathering area for cross-border families and for the general public to enjoy a scenic coastal area.
With your approach on fine tuning, will the model answer to the following prompt:
Where is located San Diego’s Friendship Park?
with something similar to:
San Diego’s Friendship Park is located on the U.S.-Mexico border.
And how would such prompt/completion line look like in your fine tuning file?
I don’t know the answers to either of those questions, but I suppose it’s probably “YES”. @Ching-Cho I’d like to use finetuning more to explore the possibilities, but based on our adjective tests (yes, all different adjectives) it seems like you can get it to respond to whatever you like. Maybe “Poem’” alone is sufficient for yours. Would like to hear what others have done with fine tuning and whether zero-prompt is possible.
For the other Q&A that @georgei raises. that sounds reasonable too. Should try it with lots of examples of Descriptive text plus Q and then A as the result. Then hopefully lots of examples later it’d mimic the format. If you did the same thing all in one prompt with even 5-6 examples in a row in davinci to test it then a few hundred to train it that would hopefully work. If you do it, please let me know, I’d like to know more about finetunning examples.
I’m actually working on this right now. I took a Word doc and sent it line by line to a JSONL file. No prompts, just fed it straight to the fine tune job. We’re talking about 100 paragraphs basically. It definitely did something, but I forgot to include stop sequences, so I’m retraining right now.
If this works we should be able to train models to be familiar with certain papers and content.
@dahifi, have you had any luck with further work on this? Similar to you (and certainly many others), I am interested in providing detailed technical papers for GPT-3 to ingest to then develop a Q&A model on a specific topic.
Not on the fine tuning end, yet, but I’ve started using gpt-index, which has a variety of index structures that you can use to ingest various data sources (file folders, documents, APIs, &c.). It uses redundant searches over these composable indexes to find the proper context to answer the prompt. I’ve only been playing with it a few days, but it’s ability to parse file folders and query against it is pretty impressive. I’m exploring it now.