3 ms·
> 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
by godelski 3y ago
> 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.