Incremental Fine-Tuning and Maintaining Conversation History

API
1

I’m new to OpenAI and am wondering if the following is possible. I would like to fine-tune a GPT-3.5 model as a starting point; let’s refer to this initial model as A1.

Then, the next day, I plan to add additional information to the model by fine-tuning it further. This process will result in a new model, named A2.

Meanwhile, my users would have interacted with model A1. If I were to replace the URL pointer to direct to A2 instead of A1, what would happen to the existing conversation history? A2 would not remember the interactions that occurred previously with A1, correct?

Is there a solution to this issue? I was hoping to be able to update and evolve the bot on a daily basis, without affecting the continuity of the chat history.

Many Thanks

2

Hey @houman ,

Welcome to the community. Let me try and answer your questions below:

I have assumed you have used an Assistant.

Technically, you are not changing the Assistant. The only thin you would do is to call the Update Assistant and update its model. You can do this at the Assistant level, or at the Run level. In either of these scenario, the Assistant will remember the previous messages.

This I believe is positive news for what you are building, and you can constantly update it.

API for updating the assistant: https://platform.openai.com/docs/api-reference/assistants/modifyAssistant
API for a new run where you can update the model_id: https://platform.openai.com/docs/api-reference/runs/createRun

1 Like
3

Thank you for your response @ idonotwritecode. I have successfully created an assistant using the GPT-4-Preview model.

However, I have encountered difficulties in establishing a consistent persona. I attempted various instructions to imbue the bot with a “mate-like” demeanor, showing interest in football (soccer) and beer, but the underlying professional character of ChatGPT often resurfaces, disrupting the intended persona.

I aimed to achieve a persona akin to those found in the Character.AI app. I wonder if the traditional text-generation (completion) approach might be more suitable for fine-tuning than the assistant model, since it allows for adjustments. However, the option to fine-tune is available only for GPT-3.5-Turbo, which is not as advanced as GPT-4-Preview. Additionally, the assistant can only be based on existing models, and I cannot use a fine-tuned model either.

I am uncertain about the best course of action moving forward.

4

I would say that it’s a bit like running a science experiment.

Run a few simple tests on all 4 systems, using the exact same query and you should be able to get to the best model (i like to do them side by side on a screen).

Do it for:
GPT 3.5 Turbo - completion
GPT 4 Turbo - completion
Assistant with GPT 4/GPT3.5