Tiny classifier lab
A hands-on lab: train a **tiny text or tabular classifier** so training/val/test and metrics become concrete.
What it is
A hands-on lab: train a tiny text or tabular classifier so training/val/test and metrics become concrete.
Why it matters
Reading about overfitting is weaker than watching a model memorize then fail a holdout set.
How it works (plain)
- Load a small labeled dataset
- Split train/val/test
- Fit a simple model (logistic regression / linear SVM / tiny net)
- Report metrics on val/test
- Inspect mistakes
Keep it small enough to run on a laptop CPU.
Everyday example
Spam vs not-spam on a few hundred emails you label yourself.
Try it
Follow the Technical lab once; then change one feature and re-check val score.
Myths
- ⚠️ Myth: Tiny labs are toy and useless.
- ✓ Reality: They teach the loop used at every scale.
- ⚠️ Myth: Highest train accuracy wins.
- ✓ Reality: Holdout metrics win arguments.
Sources
- Course 03 supervised learning; Course 21 setup
- scikit-learn user guide (cite): https://scikit-learn.org/ ↗
- fast.ai practical ethos: https://www.fast.ai/ ↗