Course 3 Week 2 Train-Dev set question

Hi, I just watch Bias and Variance with Mismatched Data Distributions. There is a new concept called Train-Dev set which I don’t get the idea well. Is it:

  • a set where you mix data from 2 different distributions while the Training set is purely from 1 distribution only
  • or both Training and Train-Dev have the same mix data, then a portion of Training set is split into a Train-Dev set

Its the second point that is right.