I have a large collection of edited documents (initial and final versions). I want to finetune a GPT model or AI assistant to act as a personalized document editor based on my editing patterns.
What’s the best approach using OpenAI tools/APIs?
2 Likes
Hi @artemreva79 and welcome to the Community!
If you are looking to fine-tune for the purpose of the model adopting a certain writing style in its responses, fine-tuning via the API is indeed a good option.
The best starting point would be the official OpenAI fine-tuning guide which is accessible here: https://platform.openai.com/docs/guides/fine-tuning
There are also worked examples in the OpenAI cookbook, such as the following:
For the customGPTs, the option for fine-tuning does not exist in the same way as it does via the API. Instead, you would largely rely on instructions and examples to get the model to adopt a certain writing style.
Let us know if you’ve got any further questions.
3 Likes
thank you @jr.2509 may i please clarify - how it difference from Assistants? I guess it is different things, but reading docs i can’t understand the difference for my case (which is building personal document editor)
1 Like
Hi! Welcome.
The difference between training a CustomGPT on the chatgpt.com ui and an Assistant via the API is the degree of control you have over the model.
CustomGPTs are paid by monthly subscription, Assistant API by data (tokens) used.
cGPTs allow you to be verbose and experiment, Assistants help you get the job done.
In your case, I would start with experimenting with your prompting in a cGPT to see how much you can accomplish. Here, you can work on refining the Instructions and Knowledge Base. You’ll want to give your model lots of examples in pairs— the original vs the edit, in addition to your overall methods of editing your work.
There will probably be a certain amount of variability in your responses from a cGPT which won’t be very useful. After the cGPT has gone as far as it can, you can take all of your training and make an Assistant—experimenting with your prompting and what not, but you can also play with the model’s native creativity (reducing it might help it edit better. 
)
After you’ve refined that as far as it will go, and if you still aren’t getting the responses you want, then you can Fine-Tune a model with multi-turn fine tuning which is what I think your case will ultimately call for.
2 Likes