2025 Contest Days

🏺 Antique

500 rows, 5 features, some labels unknown — classical sklearn beats deep learning here.

Classical MLscikit-learnSemi-supervised★★☆☆ your Week-2 checkpoint

📜The task, in plain English

A tiny tabular dataset: 500 samples, 5 features. Labels are 1, −1, or 0 — where 0 means unknown. Classify the unknowns. That's it. No images, no GPUs, no transformers.

Input	500 × 5 numeric table
Output	Labels for the unlabeled rows
Really tests	scikit-learn fluency, semi-supervised thinking, and not overfitting 500 rows

🔧The baseline you're given

Train a basic classifier on the labeled rows only, predict the rest. Limitations: throws away the structure in the unlabeled data; tiny data makes validation noisy; feature scales may be uncalibrated.

🚀Baseline vs. solution

⚠️ The baseline (what you are given)

Fits only on labeled rows
Single model, single train/val split
No use of unlabeled structure

✅ The winning approach

Cross-validation, not one split — with 500 rows, a single split lies to you
Semi-supervised tricks: sklearn.semi_supervised label propagation / self-training (predict unknowns, add confident ones to training, repeat)
Compare several cheap models (logistic regression, random forest, gradient boosting, SVM) — 5 minutes each
Scale features; plot them — with 5 features you can literally LOOK at the data

🧒Explain it like I'm brand new

Why this task exists: to catch people who reach for deep learning on 500 rows. Small tabular data is classical-ML territory: fit/predict in scikit-learn, cross-validate honestly, and exploit the unlabeled rows (their positions in feature space are information even without labels — points in the same cluster probably share a label; that intuition is "label propagation").

The muscle: the sklearn workflow — fit → predict → cross_val_score — which you can run in your head. This is the Week-2 checkpoint of your 30-day plan for a reason.

💬The Gemma 4 playthrough (2000-token limit)

Chat 1 · Fast model bake-off

YOU

sklearn: X (labeled rows, shape (n,5)), y in {-1,1}. Write ONLY code to compare LogisticRegression, RandomForest, GradientBoosting, and SVC with 5-fold cross_val_score (accuracy), with StandardScaler in a Pipeline, printing mean±std per model. Max 25 lines.

Chat 2 · Self-training

YOU

Now ONLY code for sklearn SelfTrainingClassifier wrapping the best model, using X_all where unknown labels are -1... wait, my unknowns are 0 and real labels are 1/-1. Handle relabeling so sklearn's convention (-1 = unlabeled) is respected. Max 20 lines.

That mid-prompt correction is realistic — you caught a convention clash (sklearn uses -1 for 'unlabeled', the task uses 0). Noticing such traps is exactly the code-literacy the contest tests.

🎯Takeaways & what Day 1 might do with this

Pattern family: classical ML on tiny data. Deep learning is a hammer; not everything is a nail.
Trust only cross-validation on small data — single-split scores swing wildly.