Can a fine-tuned GPT model handle dynamic information during inference, and is it effective to use markers (e.g., brackets) in training data to indicate variable parts for better generalization?

API
1

Hello!

I’m currently working on fine-tuning a GPT model and have a couple of questions:

  1. Can I provide a dynamic system prompt to a fine-tuned model during inference? For example:

    "Never forget your name is {salesperson_name}. You work as a {salesperson_role}.  
You work at a company named {company_name}. {company_name}'s business is: {company_business}.  
Company values are: {company_values}.  
You are contacting a potential prospect to {conversation_purpose}.  
Your means of contacting the prospect is {conversation_type}."

    If I dynamically replace the placeholders in {} with actual values, will the fine-tuned model understand and handle the changes effectively?

  2. During fine-tuning, can I include placeholders like {} in the system prompt in training data to indicate that these parts will vary dynamically at runtime? Will this help the model generalize better for scenarios involving changing input data?

  3. A critical requirement for my chatbot is maintaining conversation history, which the GPT model does not do by default when accessed via the API. How should I structure inputs for the fine-tuned model to include and effectively process conversation history?

  4. Finally, when interacting with a fine-tuned model via the API, can I pass additional information beyond the system prompt and conversation history? If so, what is the best way to include this supplementary information to ensure optimal results?

I’d appreciate any insights or advice from those with experience in fine-tuning GPT models. Thank you!

1 Like
2

I think before rushing to finetuning you can use a RAG maybe even with a simple rdbms behind and not just a vectordb.

postgres + pgvector seems tempting although it has it’s limitations…

1 Like
3

This might help you. ChatGPT - 3Dเจวิส

4

I understand I need RAFT) but I don’t understand is it possible to do with Chat GPT from OpenAI? and if yes please share link how to do or some example