learning curve for Neural network / Machine learning

I understand the idea of learning curve and how to solve problems based on its information.

https://www.youtube.com/watch?v=g4XluwGYPaA

QUESTION: If we take the MNIST task with 60k training set + 10k test set.

I will vary my training set to evaluate, but how big should my test set to be for this. E.g.. if I only use 2000k training set should my test set can still be 10k?