Replies: 6 comments
-
interesting idea, like that one game show i'm not gonna name. I can see the potential benefits. |
Beta Was this translation helpful? Give feedback.
-
By prompt, do you mean the prompt that would generate the reply? I wonder what we could expect to see from that. I'm assuming the user would have to provide something that would work as an initial prompt, because "Now give me a list of these places." wouldn't be a very effective example. Then again, the assistant could be making a reference to an earlier statement, in which case the correct prompt would be a user reply - unless you limit the assistant messages to the first ones of each tree only. |
Beta Was this translation helpful? Give feedback.
-
Yes |
Beta Was this translation helpful? Give feedback.
-
This could be used as a way to train and tune doc2query as well - there a somewhat related issue in this project with #645. Could also be used to connect a given response to more prompts and thereby increasing the dataset for the reward model to train on. |
Beta Was this translation helpful? Give feedback.
-
This is actually a brilliant idea @erkinalp. This will create diversity for the data. You would then be creating a new sister subtree tho. So this could complicated things. I will leave for the webteam to decide if this can be implemented. i adding appropriate labels to let that team know. |
Beta Was this translation helpful? Give feedback.
-
Maybe the trees created from this should be in a different dataset altogether? That'd make things easier, but I don't know if there's a benefit in putting the predicted prompts in the same dataset. |
Beta Was this translation helpful? Give feedback.
-
This training scenario might be useful to train latent diffusor (http://arxiv.org/abs/2212.09462) and combined diffusor+GAN architectures of generative models.
Beta Was this translation helpful? Give feedback.
All reactions