I’ve encountered some issues while attempting to fine-tune the GPT-4o-2024-08-06 model using image data.
First, according to OpenAI’s documentation, GPT-4o models should support image fine-tuning, and I understand there are 1M free training tokens available daily until October 31, 2024. However, my fine-tuning job resulted in a billing error:
“Creating this fine-tuning job would exceed your hard limit. Please check your plan and billing details. Cost of job: USD 9.79. Quota remaining: USD 9.43.”
This was my first attempt at fine-tuning today, and I am unsure why it was charged, given the free token offer.
Additionally, I also faced an error related to image content in my training file. Some images were skipped with the message:
“Contains people. These examples will not be used for training.”
I’ve followed the formatting guidelines but am still facing issues.
Lastly, I also tried using the GPT-4o-mini-2024-07-18 model as mentioned in the API reference, but the fine-tuning job failed with the error:
“Invalid file format for this model. Images are not supported for gpt-4o-mini-2024-07-18.”
Could you please help clarify:
Why the free training token limit is not being applied?
How to address the image content issues?
Whether the GPT-4o-mini model supports image inputs as per the documentation?
You can find more details about the fine-tuning process here.
Have you considered the number of epochs in your calculation of total tokens? This could potentially explain the deviation.
(base training cost per 1M input tokens ÷ 1M) × number of tokens in the input file × number of epochs trained
I could be wrong but my understanding based on this post is that currently only gpt-4o models can be fine-tuned with images.
Yes, images with people and faces are excluded as per the moderation policy. This is also documented here.
I hope that addresses some of the issues. If you have more questions, then you might want to add them to this thread where OpenAI staff has been pretty active in responding.
gpt-4o-mini does not support image fine tuning, I made the same mistake! there’s a specific gpt-4o model you need to use, the docs make it a bit confusing, but there is a blog page that at the end says the right model to use
Thank you very much for your responses. I completed the training, and I can confirm that it was free; the message I received was merely a reminder to ensure I had sufficient credits in my account. The fine-tuning with the GPT-4o model worked well; I understand that it is not yet available for the mini version.
