Skip to main content

Lesson 10 · 12 min

Capstone — diagnose a training run

You're handed a broken run. What's wrong, and what do you check first?

The brief

A teammate says "I trained a classifier and the model is bad. Help me debug." You see:

  • Train accuracy: 52% (chance is 50%)
  • Val accuracy: 51%
  • Training ran for 50 epochs
  • Loss curve: roughly flat the whole way through