I wan’t to use GPT-3 with very specific data f. e. about “Falaffel”. Therefore I want to submit hundreds of pages about the history of falaffel, recipes, … and then I want to generate content based of the input I gave and I want an option to create more content later on my submitted data. Is this possible?
2 Likes
Yes, this is possible, but your results might vary with just completion.
Here’s some links on fine-tuning…
You might also look into embeddings endpoint…
Hope this helps.
2 Likes
What I don’t understand about either example is how to get special knowledge into the fintuned model. After all, the examples show how to set a sample query with a perfect result. Where does the system get the “basic knowledge” about the area in the first place, from which you can then generate/query specific content?
A concrete example. I want to develop quiz questions about a city. I have documents about the history of the city, inhabitants, companies of the city, personalities of the city, … but the city is not relevant enough for the standard models to have information about it.
So as an example:
Generate 2 Quiz-Questions with 1 correct and 5 wrong answers about Berlin.
-
What is the capital of Germany?
A. Berlin
B. Munich
C. Cologne
D. Hamburg
E. Frankfurt
-
In which year did Berlin become the capital of Germany?
A. 1871
B. 1789
C. 1990
D. 1945
E. 1949
is great but
Generate 2 Quiz-Questions with 1 correct and 5 wrong answers about the village Breuningsweiler.
Question 1: What is the population of Breuningsweiler?
A. 10.000
B. 5.000
C. 1.000
D. 20.000
E. 50.000
F. 500
Correct Answer: B. 5.000
Question 2: How many schools are in Breuningsweiler?
A. 4
B. 2
C. 10
D. 6
E. 8
F. 0
Correct Answer: B. 2
Is completely wrong because GTP-3 won’t have data about that <1.000 residents village in south germany.
Would I have to create a promt on each information about the city OR is there any way to provide a lot of plain text (like from the wikipedia website about breuningweiler, the tourist information website, …) to give the system an information base?
You can fine-tune without a prompt, but GPT-3 is known to hallucinate at this number of parameters, so it’s not 100% great at being an expert on a specific domain.
Have you read up on fine-tuning and embedding? There’s a few posts on the forum here that might be helpful too.
Good luck.,
2 Likes