Why Isn't It Possible to Finetune ChatGPT-4o?

I’m curious about the finetuning capabilities of ChatGPT models with vision, particularly the new ChatGPT-4o. This model seems to perform better on vision tasks than GPT-4 V from my limited testing.

However, I’ve noticed that the latest finetuneable model available is GPT-4-0613, which only includes data up until 2021. Given the advancements and potential improvements in more recent models like ChatGPT-4o, it would be incredibly beneficial to have the ability to finetune these newer versions.

Could someone shed light on why it’s not possible to finetune ChatGPT-4o? Are there any plans to enable finetuning for these advanced models in the future?

Thanks in advance for any insights!

1 Like

Hello,

ChatGPT-4o, the newly released model with integrated visual capabilities, indeed shows superiority in visual tasks. However, there are several reasons why ChatGPT-4o is currently not available for fine-tuning:

  1. Technical Complexity: Models that integrate visual capabilities are more complex than text-only models, and developing frameworks and tools suitable for fine-tuning takes time.
  2. Testing and Stability: New models require extensive testing to ensure stability and reliability before officially launching fine-tuning capabilities.
  3. Policy and Legal Considerations: As model capabilities increase, the regulatory and ethical requirements for their use may also rise.
  4. Resources and Support: Providing technical support and maintaining relevant APIs and tools require substantial resources.

Although it is currently not possible to fine-tune ChatGPT-4o, the possibility of opening it up for fine-tuning in the future exists. It is advisable to keep an eye on official announcements and updates, and in the meantime, you can continue exploring existing fine-tunable models to meet your needs.

I hope this helps your understanding!

Thank you for sharing your insights! I hope that GPT-4o will be finetunable soon though. It performs drastically better (without finetuning) on the task I need it for and previous models just don’t suffice.

I completely understand and share your anticipation for the fine-tuning capabilities of GPT-4o. I, too, hope that in the future, we can apply fine-tuning on even more powerful models to fully leverage their potential. Let’s look forward to technological advancements and the rollout of new features.