Let’s say I’ve fine -tuned a given chatppt api model (e.g. chatgpt 4.x) with my own documentation. What happens when there is a new model available (e.g chatgpt 4.y). Do i have to fine tune the api again?
1 Like
Hi @marta7 and welcome to the Community!
Couple of points on this:
-
OpenAI’s fine-tuning endpoint is not designed to provide models with new knowledge. Instead, fine-tuning can serve to get the model to respond in a certain style, format or to get to perform tasks in a specific way. In your case, you want to look into a RAG solution. You should have a look at the following guide which explains the difference between the two options.
-
To nonetheless address your core question: if you were to fine-tune a model, then typically these can be accessed for a very long time-frame, often even after a base model was deprecated and replaced by a newer version. Ultimately, you will have to re-do the fine-tuning as models change and you want to take advantage of improved model capabilities.
4 Likes
ChatGPT is a name specifically for OpenAI’s own end-user ChatBot.
AI models on the API have specific underlying versioned names, such as ft:gpt-3.5-turbo-0613:organization::idnumber that are now based on a release date.
Let’s say I performed a fine-tuning on GPT-3 model “davinci”. That is now shut off. Any fine tuning that was performed is now gone along with the model being inaccessible (although some received credits against some of the past fine-tuning performed).
How about that gpt-3.5-turbo-0613 model, just shut off last month? The fine-tune is still working for now, although not returned by the model listing endpoint - so you better have written down the ID. When it will actually stop working is currently unclear.
The fine-tune model you make today will continue working past some new versions being introduced, and is only trained against the one specific model version, but there will come a point where the switch is ultimately thrown on that as an older model also.
Read more about deprecations and shut-off:
2 Likes