Hi,
I want to retrain a fine-tuned model. However, I’m unable to do so. I get the following error:
{
"error": {
"message": "Model <model_name> is not available for fine-tuning or does not exist.",
"type": "invalid_request_error",
"param": null,
"code": "model_not_available"
}
}
This is very strange because I have done the same thing in the past.
Knowledge around any changes that are causing this would be very helpful since I am unable to find anything in their documentation.
I’d wonder:
Are you using an API key that can list the ft: model name in the models endpoint?
Is this a base model that is still available for fine-tuning for you, such as being listed in the drop-down of the platform site dashboard UI? Are the target fine-tunings appearing there too as base models? Note the org->project selected at upper left.
Yes to the first 2 questions.
Can you clarify the last one?
In the drop-down for selecting a base model in the platform site for fine-tuning, you have or should have both the ‘plain’ model such as gpt-4o-2024-08-06, and then also, scrolling down, the model listing you can pick from should also then include prior fine tuning models, with names like ft:gpt-4o-2024-08-06:organization:prefix:3e1443 or similar.
If the fine tuning models that you can currently employ for inference are not listed there to start a job in the provided user interface either, then one might guess that something has gone wrong with the project settings or its rights to use models or the fine tuning endpoint. First check project limits and reset all models to block->“non checked”. Then make an API key also that doesn’t have any endpoint blocking.
Otherwise, if there is no listing of fine tuning models in the ‘you can continue fine tuning’ UI, it is likely the fault that is going to block you running that by API also. If it is a current model that the fine tune is based on, than any corruption preventing access would need to be fixed by OpenAI.
Or, I suppose, just start again and make a single combined fine-tuning with all the examples. OpenAI automatically adapts the learning strength to the file size, and doesn’t consider that you’d want equal weight with existing training in continuations (thus you needing to reduce the learning rate multiplier and only run a minimal n_epochs to not degrade what’s already reinforced).
Yes, I can see my prior fine-tuned model in the base model selection UI. However, even selecting it there and then starting a fine-tuning run gives the same error.
I don’t think a single combined fine-tuning will work for me. We have tried different variations in the past and multiple fine-tunings work the best.
I have the same issue. Previously, first running a supervised training run followed by DPO was not just possible but is still the recommended procedure. Now this does not seem to be available anymore.
I think @_j is suggesting a possible workaround related to project permissions.
You could try creating a new project, adding the fine-tuned model, and then continuing the training.
This might help narrow down the cause of the issue and could get you unblocked for now.
Your proposal is unlikely to succeed.
Fine-tuning models are now scoped to a project, and also have a nasty concept of “created by me” pinned to them (just like full user names now pegged throughout the platform site and usage.)
Switch project ID on existing fine-tuning view in the UI, that had a listing and one selected:
I’ll message staff to take a look.
Thanks for checking!
It may be that they are all dumped into “default project”, as it makes no sense that fine-tuning jobs created by a user key - an owner ID - disappear when changing projects. I suspect that forum regulars have been fine-tuning models longer than the people pushing these nonsense changes onto API assets that have been paid for have been at OpenAI.
Hi,
I’m experiencing exactly the same issue.
I’m trying to fine-tune a model that has already been fine-tuned (using GPT-4o or GPT-4o-mini), but I keep getting an error both via API and from the web interface.
Problem details:
I fine-tune a model based on gpt-4o-mini-2024-07-18.
The fine-tuning job completes successfully and the resulting model works as expected.
When I try to fine-tune that fine-tuned model again (via API), I get this error:
Model ft:gpt-4o-mini-2024-07-18:<org>:<model-name>:<id> is not available for fine-tuning or does not exist.
I’ve waited several days, but the issue persists.
From the web interface, it lets me select the model and upload the .jsonl file, but when I click Create, I get this error:
Error creating job: Model ft:gpt-4o-2024-08-06:<org>:<model-name>:<id> is not available for fine-tuning or does not exist.
Interestingly, I can fine-tune older models (also based on gpt-4o-2024-08-06 or gpt-4o-mini-2024-07-18) that were created a few months ago.
Yes, the same thing is happening with me
This issue is now resolved. You can now retrain newly fine-tuned models. Thanks everyone!
@ssm can you share how you made it work?
the problem is still occurring for us.
You must be trying to retrain a model that was fine-tuned a few days back. As I mentioned, try retraining models that are newly fine-tuned(that is, fine-tuned on or after today).
Exactly, that matches my experience. Models I fine-tuned a few days ago still can’t be retrained, but I just trained a new model today and was able to fine-tune it again without any issues. I hope it will be possible to fix the older models that are currently blocked.