Hello.
I’m interested in fine-tuning, and I’m actually trying to do some tuning.
However, the official documentation alone was not very specific, and newbies like me didn’t understand good public procedures.
Do you have any other reference sites?
This might be the wrong subforum - there’s the API forum for something like this.
I wrote a short guide at the bottom here with some sample files: GitHub - smuzani/openai-samples
Let me know if you get stuck on something; I assumed some basic python knowledge.
However, it might not be that useful to learn unless you have a lot of data and need the added accuracy. Fine tuning is probably the most expensive option today, now that there’s the gpt-3.5-turbo model available. There’s also options like embeddings which have become much cheaper.
3 Likes
Mr. smuzani
Thank you for teaching me carefully.
I will take a close look at the information you have given me.
Giving me options was very helpful.
1 Like
Hi.
Certainly fine tuning was too expensive.
Does the embedded option you told me refer to parameters when calling the API?
Please let me know if there is any documentation to help me understand.
1 Like
Oh, sorry I should have given more details.
If you’re trying to guide it to have a certain tone, consider using the chat model to give it more examples of how it should respond, via the assistant parameter.
If you’re trying to give it a knowledge base, embeddings can work better. There’s some examples of doing search for a doc: OpenAI Platform
Here’s an even better example of a type of “chat with document”: Chat with Document(s) using OpenAI ChatGPT API and Text Embedding | by Sung Kim | Dev Genius
thank you!
I didn’t understand everything you taught me, but I’ll try to learn so that I can understand it from now on.
1 Like
Hi Mr. smuzani
I read the document you gave me.
Among them, a variable called chat_history suddenly appears. What is this?
What it says on the tin.
You should check out the hugginface courses, the one on NLP and transformers should prove illustrative.
Fine tuning even training shouldn’t prove that expensive depending on the model.
If you are interested in other models and local llms, hugginface is where its at.
1 Like
Embeddings are just a clever way to get context from your database into your prompt (with semantic search)
User input-> Find relevant context with embeddings-> Place context into prompt-> Get chat completion
This doesn’t train the underlying model but acts as a smarter way of prompt engineering.
2 Likes
Mr. Anaryl
thank you!
I’ll take a look at that page!!
1 Like