#top-p

Question 1: Single choice

A language model is used for two tasks: auto-generating legal contracts and writing poetry. Which choice of decoding strategy fits the two tasks best?

A high temperature for the contracts so the wording varies, greedy decoding for the poems so they stay coherent1
Greedy decoding for both, since it always produces the most accurate text2
Greedy decoding or a low temperature for the contracts, top-p sampling or a higher temperature for the poems3
A high temperature for both, since it makes the model's output more accurate4

Timed test