The Importance of Labeling Answers

Some of these papers can get long and there's a limited window to type into chat GPT or to send to the API. Did you hit that ceiling? For sure when we originally tried copy pasting the entire paper and we weren't able to do so. And performance wise now you do have some data you can look at it's not necessarily where you have to manually review the text We can predict through the yes or no answers that the model comes up with match what the humans gave. The accuracy jumps up to 92.8% which is higher than the human authors but we caution against using this 92.8%.