Can we fine-tune instruct-series GPT3 Models?


I would like to know if it’s possible to fine-tune instruct-series GPT3 Models. Are there any drawbacks of doing so than fine-tuning general models.

Thank You!

I don’t believe it is possible yet but it would be great if we could.

Hi @SecMovPuz @akarshg! what would you be hoping to get from fine-tuning instruct models? You should be able to get many of the benefits from fine-tuning a base model.


I think especially with the new updated information and the different way training can be formatted for the instruct series it would at the very least be easier to create fine tune documents for it. Also having the 4000 token limit would be great.

Hi @luke. I am confused by your answer. Aren’t the default engines the instruct series? And I believe those can be fine-tuned, correct? More generally, I think there is confusion circulating about the updated engines/models; the new token limits which apply in some but not all circumstances; and which models/engines can be fine-tuned and which cannot. I think the documentation (and the examples) could really use an update. It would also be great if sections in the documentation were tagged “last updated xx/xx/20xx” and if there was a dedicated page listing technical amendments in reverse chronological order. The announcements page doesn’t seem comprehensive/reliable and I get the sense that people are relying on word-of-mouth to learn about changes. Thanks for providing this forum where we can give feedback. Best, Leslie


Agreed! Any examples in particular that would be helpful?

This is helpful feedback! The original models are available for fine-tuning right now. We’re working on making the latest text-davinci-002 available for fine-tuning too.

1 Like

Probably making the token limits with the new model more clear.

Also having new features clearly explained, possibly in a separate page would be good.

More examples of newer models and more comparisons to older models would be nice. I noticed that occasionally for some tasks the original davinci can work better then instruct. I think the main thing is that there are features that are not clearly stated or not stated at all in the documentation.

Anyways keep up the good work. I’m glad you guys are open to suggestions and care about the community.


I didn’t realize that if I’m using text-davinci-002 I can’t fine tune. Grrr…really need to clarify all this in the documentation, perhaps using a chart with these headings:
->Model Name
->Corresponding Engine(s)
->Token Limit
->Is Fine-Tuning Possible? (y/n)
->Link to example colab file showing usage with a simple jsonl file containing a few lines of data in the required format