Feb 7, 2021
How much training data do you need to get measurable improvements when you finetune? Do something like 20 to 50 question/annotation pairs suffice, or do you only see improvements on much higher numbers in the hundreds or thousands?
How much training data do you need to get measurable improvements when you finetune? Do something like 20 to 50 question/annotation pairs suffice, or do you only see improvements on much higher numbers in the hundreds or thousands?
