Pointwise V-usable information (PVI) excels in many
#NLProc
tasks. But fine-tuning
#LLMs
with it is very time-consuming π Is in-context PVI the necessary next step?
Yes! πCheck out our empirical analysis accepted at
#EMNLP2023
β and this 𧡠(1/7)
π
Pointwise V-usable information (PVI) is a recently proposed metric for measuring the hardness of individual instances. It is estimated by fine-tuning supervised models. (2/π§΅)
#EMNLP2023
In our paper we show that in-context PVI exhibits similar characteristics to the original PVI but is more time-efficient. The reason is that it requires only a few exemplars and does not need fine-tuning. (3/π§΅)
#EMNLP2023
Our findings show a lower prediction accuracy for low in-context PVI (see in the π¦ box); and a higher average in-context PVI for correct predictions than incorrect (see in the π₯ box). This matches what we see in the original PVI estimates. (4/π§΅)
#EMNLP2023
Major insight : in-context PVI estimates are more consistent across similar models (e.g., models that have similar architecture or similar training data. (5/π§΅)
#EMNLP2023
We also show that, comparable to the original PVI, the in-context PVI threshold at which instances start being predicted incorrectly is similar across datasets. However, in-context PVI estimates made by smaller models are much noisier than those made by larger models. (6/π§΅)
Besides, in-context PVI estimates can be used to identify mislabeled instances. This is a very practical feature and demonstrates the reliability of the in-context PVI. (7/π§΅)
#EMNLP2023