Applied Case — The Honest Tabular Model, End to End

P5.applied-cases.01 · Audience: guest, it-ml, language-pro · Prerequisites: Model Evaluation, Selection & Interpretability, Evaluation Metrics for NLP

LLM & Generation Evaluation Applied Case — Clustering & Segmentation

LLM grading (this page):Real gradingOff — no LLM calls

Knowing every technique in the catalogue does not by itself produce a trustworthy model: the number you report at the end is only honest if the steps are chained in the right order, with the test data locked away before anyone looks at it. This module runs everything the ML Methods and Evaluation tracks teach once, in that order, on real data — the guided walk builds a species classifier on the Palmer penguins study table with every number shown, and the practice ladder then makes you rebuild it rung by rung in the kernel — the server-side workspace that runs and grades your code — against real files, real missing values and resource-metered grading. It ends in the platform's first Project: the same arc on 32,561 census rows with genuine missingness and class imbalance, graded by hidden tests and reviewed against a design rubric.

Step 1 / 8 — Frame the task — and meet the data

Task: predict a penguin’s species(Adelie / Gentoo / Chinstrap) from its measurements — a supervised, multi-class classification problem, exactly the shape catalogued in Supervised Learning Algorithms. The study table has 344 labelled rows and 8 columns; some measurements are missing, and sex is missing on 11 rows.

Column	Type	Missing
`species` (target)	categorical, 3 classes	0
`island`	categorical, 3 values	0
`bill_length_mm`, `bill_depth_mm`, `flipper_length_mm`, `body_mass_g`	numeric	2 each
`sex`	categorical, 2 values	11

Class mix: Adelie 44.2 %, Gentoo 36.0 %, Chinstrap 19.8 %. Not balanced — every later choice (stratified split, baseline, metric reading) has to respect that.

Self-check

Why does framing come before any code — what could go wrong if you decided the task after exploring the table?

📚 Go Further

Type	Resource
In-app	P5.ml-methods.06 — Model Evaluation, Selection & Interpretability — the splitting, CV, leakage and selection theory this case exercises
In-app	P5.evaluation.01 — Evaluation Metrics — precision/recall/F1, ROC/AUC and calibration, with the interactive threshold explorer
In-app	P5.applied-cases.04 — The Leakage Clinic — what happens when the discipline above is skipped, three ways
External	scikit-learn user guide — Pipelines and composite estimators; Common pitfalls: data leakage
External	Mitchell et al. (2019), Model Cards for Model Reporting — the template step 8 abbreviates

Ask the mentor about this module

Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.

Images, PDF or text. Kept on this device only.

Keeping your files on this device

Keep a copy of the files I attach, in this browser

Off by default. The mentor always gets your file; this only decides whether your own copy stays here. Copies live in this browser only - they do not follow you to another device, and clearing site data removes them.

Ctrl/Cmd + Enter to send

🎓 Practice ladder

5 graded rungs · ~97 min

Now build it yourself, in the kernel — real files under /data, real missing values, resource-metered grading. Four guided rungs rebuild the penguins walk; the finale is the platform's first Project: the full arc on the UCI Adult census table, graded by hidden tests and reviewed against the four-criterion design rubric shown in the workspace.

Rung 1 — Split before you look

Loading exercise…

Rung 2 — Baselines before models

Loading exercise…

Rung 3 — Preprocessing belongs in the pipeline

Loading exercise…

Rung 4 — Compare models with cross-validation

Loading exercise…

Rung 5 — Project: the honest income model

Loading exercise…

✅ Key takeaways

The order is the method: frame → split → explore (train only) → baseline → pipeline → cross-validate → threshold → error analysis. Every stage exists to protect the honesty of the final number.
Split before you look, stratified on the target — 275/69 on penguins — and spend the test set exactly once.
Preprocessing lives inside the Pipeline: imputers and encoders fit on training rows only, and prediction time replays the identical transformation (leakage-proof and serving-proof at once).
Model selection runs on cross-validated training folds (logistic 0.9927 vs tree 0.9527); the test split votes zero times in any decision.
On imbalanced data the threshold is a product decision: sweeping the precision-recall curve on the census model lifts positive-class F1 from 0.6734 to 0.6941 over the 0.5 default.
The deliverable is model + card: data, split protocol, metrics, intended use, explicit non-uses — the artefact P13's serving-and-drift case picks up.

Next: P5.applied-cases.04 — The Leakage Clinic: three notebooks that score brilliantly and would fail in production; find each leak, fix it, and reproduce the honest number.

Try it yourself

A scratch console for this page's ideas — ungraded, nothing you run here is recorded.

Scratch console

A scratch console with the scientific stack (pandas, numpy, scikit-learn). Runs on the server — no network, resource-limited and measured.

Output appears here.

My notes on this module

Loading your notes...

Where next?

Continue

Applied Case — Clustering & Segmentation →

Applied Cases · P5

Later in Applied Cases

This module unlocks

Go up a level