Re:What we learned from using GPT for 500k+ classifications
by mistymountains on Tuesday July 11, 2023 @07:34PM
(#36688589)
The training classes in this case are words in the vocabulary in the context of a sentence.
Re:What we learned from using GPT for 500k+ classifications
by godelski on Tuesday July 11, 2023 @07:47PM
(#36688684)
> It is zero shot. The llm is trained to generate next token.
I'm going to refer you to the sibling comments as they stated similar things and I answered them in depth and do not wish to repeat myself.
But to summarize:
Zero-shot := Goal of f:X → Z but train f':X → Y, where Y ⊂ Z. We test on A⊂Z, where A⊄Y (sometimes definition is A ∩ Y = {∅}, but I'm not being as strict)
> but that the model was not trained on this particular ML task and/or classes.
I'm going to need explicit clarification as to this. Explicitly or implicitly? See sibling comments and note about likelihood and density estimators w.r.t. classification.
Zero shot is defined as being able to output predictions for classes they were not trained on.
It doesn't mean the input data can't be in ml task domain but that the model was not trained on this particular ML task and/or classes.