-
-
Notifications
You must be signed in to change notification settings - Fork 22
Description
Sometimes editing prior messages, including the output of the model is useful for pushing it in the right direction.
This feature is available in the Playground mode.
By default, in this project the ChatGPT chat log buffer is read-only. However, disabling this and editing the chat does not change the data that is passed to the model. I assume that this is because the data in the buffer is not literally the raw data.
You can see an example here where I edit the model's output, and would expect it to give me the capital of Sweden.
Essentially, I think a useful feature would be for the chat buffer to be bidirectional.
user: Hello there
[assistant]: General Kenobi!
[assistant]: I apologize for the reference. It's a line from Star Wars: Episode
III - Revenge of the Sith. In the movie, when Obi-Wan Kenobi encounters General
rievous, he greets him with "Hello there."user: I see...
[assistant]: I'm sorry for providing a confusing answer, can I help you with anything todayediting the assistant output
[assistant]: I am about to tell you the capital of Sweden
user: So, what is it?
[assistant]: I'm sorry, I'm an AI language model and I'm not sure what
specifically you are referring to. Could you please provide more context or
clarify your question? I'll do my best to assist you.