Hi I am looking for some help solving my issue with my fine tuned model of gpt4o mini.
The fine tune job was successful, and the model works pretty much how I intended. I can chat with it on the open ai API chat playground and use the endpoint to communicate with it directly through postman etc.
However, after I create an assistant either using my backend or directly in the assistant playground using my fine-tuned model the run fails immediately with any prompt. It responds with Run failed Sorry, something went wrong.
{
“response”: “Sorry, something went wrong.”,
“error”: “server_error”
}
Like I said the model works fine when directly prompting it, and the assistant works with a base gpt4o mini model with the same instruction set and any prompts.
Any insight would be appreciated.
I’ve had people say that gpt4o mini fine tuned models dont work with the assistant api but I don’t see why I would be able to select it as a model in the playground then.
But the openAI Support said - “Based on the latest information, fine-tuned models of GPT-4o mini are indeed supported in the OpenAI Assistants API”.
Attached is a picture of the playground errors when trying to use my fine tuned model. I have tried both simple prompts and prompts more in lined with what my fine tune model would expect. I have also tried the assistant with and without the code interpreter and file search tools.
I’ve just completed a test with a fine-tuned gpt-4o-mini model on my end and was able to successfully use it with an Assistant in the playground.
Typically, a server error indicates a temporary issue on OpenAI’s end. I would give it another try today and if the error persists, reach out directly to OpenAI support via the chat function for them to have a look at your case.
Interesting. Thanks for sharing that information - that would be new to me too and having just looked over the docs and FAQ again, I also could not find no explicit information in that regard.
I’m in Tier 4 with my main account but will see if I have an opportunity to retry on my end with another account in a lower Tier.
Hi again @ryanjbellmore - the Tier should not matter. I was just now successfully able to use a fine-tuned gpt-4o-mini in the Assistant’s playground with an account in Tier 1.
I hope it’s just a transient error on your end that resolves itself.
Thanks for taking the time to investigate it further, much appreciated.
I did think it would be weird that using a fine-tuned model would require a specific usage tier and make no mention of it anywhere.
This what the response I got from the OpenAI Support -
“The issue you’re encountering, where you receive the error {“response”: “Sorry, something went wrong.”, “error”: “server_error”}, seems to be related to your usage tier. As mentioned in the response you shared, fine-tuning for GPT-4o mini is restricted to users in higher usage tiers, specifically tiers 4 or 5. If your account is not in these tiers, this could be causing the problem you’re experiencing when attempting to use the fine-tuned model with any assistant.”
Unfortunately, I am still receiving the same error still and I tried to retrain the model and use a new API key as well.
Ah I see. What they are referring to is the fact that initially fine-tuning of gpt-4o-mini was restricted to developers under Tier 4 and 5. This limitation was however recently lifted and now developers across all Tiers can benefit from the option to fine-tune.
In other words, they misunderstood your problem
If the problem persists, perhaps try another outreach and rephrase the issue a bit so it’s clearer to support.
