Hello Model
← Training basics

Lesson 3 of 8

Train, validation and test splits

In plain words: Split your data so you can check the model on examples it has never seen.
  • Train (~70–80%): the model learns from this.
  • Validation (~10–15%): you use this to compare models and tune settings.
  • Test (~10–15%): touch it once, at the end, for an honest final score.
  • If your data has dates, split by time so the test set is 'the future'.
  • Watch for data leakage: duplicates or features that secretly contain the answer.
Tip If your test score is suspiciously good, look for leakage before celebrating.