When building my custom GPT, I’ve noticed that it’s very easy to get the desired response by providing feedback.
I guess GPT Builder uses Reinforcement Learning behind the scenes.
I was wondering if there is a similar feature in the assistants api, which allows the user to provide feedback on a thread.
If not, I’d like to add this as a feature request.
You need to opt-in for this to happen.
While the voting system for ChatGPT and the custom GPTs is used to adjust the weights of the model, anything being send to the API is not used for training as a baseline rule.
This has been a recurring request from developer’s who handle sensitive, internal and proprietary information.
Thus it is not possible to implement a similar mechanism for the Assistants API.
But, you can specifically opt-in using this form:
There is no API for sending your satisfaction or disapproval of a generation result to OpenAI.
It is expected that your application will be quite different than the baselines.
It actually would be nice to have an API-tuned model that obeys no “ChatGPT” training, that is competent at a vast variety of instructed tasks, but instead we get more and more models that only have been trained by users of ChatGPT.