Diagnosing and tuning: is the number solid, would more data help, is there better?

Step 1: A model worth diagnosing10 min

1. New project PW12 Diagnostics, add Prix immobilier Californie.
2. Train Modèle de référence (médiane) (dummy_reg) on price. Experiment label diagnostic.
3. Train Régression linéaire (linreg).
4. Train Forêt aléatoire (rf_reg) with default settings (100 trees, max depth 10).

Model	R²	RMSE	MAE
Baseline (median)	-0.0540	1.1809	0.8813
Linear regression	0.5911	0.7356	0.5297
Random forest	0.7795	0.5401	0.3646

Three numbers. The rest of the session asks whether the third one is worth anything.

Step 2: Learning curve: would more data help?15 min

1. Open the forest results, scroll to the Diagnostics tab.
2. Learning curve → Compute. The same model is retrained on five growing samples, three folds each. On this dataset it takes about eighteen seconds and a progress bar tells you which fold it is on. You can cancel.
3. Read the two curves and, under them, the sentence the studio computes.

The four possible readings, and what each one means for your next hour:

Reading	What it means	What you do
The gap stays wide (0.10 or more)	the model memorises rather than generalises	simplify the model, or get more data
Both curves stay low, together	the model is too simple for this data	more data will not help; change the model or the features
Validation has plateaued	more data would add almost nothing	change the model, not the dataset
Validation still rising at the largest size	adding data should still gain score	go and get more data

Step 3: Cross-validation: is the number solid?12 min

1. Cross-validation, k = 5, Compute. About twelve seconds on this dataset.
2. Read the mean, the standard deviation across folds, and the sentence.

Now use it properly. Train a Gradient Boosting (gb_reg) and cross-validate it too. Then answer, in writing:

• Is the difference between your two models bigger than the standard deviation of either of them?
• If it is not, what have you actually demonstrated?

Step 4: Validation curve: where overfitting starts13 min

1. Validation curve. The hyperparameter selector is built from the registry: for Forêt aléatoire (rf_reg) it offers Arbres (n_estimators, 10 to 500) and Profondeur max (max_depth, 1 to 40).
2. Choose Profondeur max and compute. Five to eight values between the registry bounds.
3. Read where the training curve keeps rising while the validation curve turns down. That point is the entry into overfitting, and the studio names it in a sentence under the chart.
4. Then try Arbres. Notice that the curve is nearly flat.

Write down the answer to: which of the two hyperparameters is worth your attention on this dataset, and how much score is at stake?

Step 5: Search, and apply without losing anything15 min

1. Search, kind Grid, pick both hyperparameters. Up to 25 combinations. About 25 seconds here, and the note will tell you if the axis was shortened or the rows sampled.
2. The result is a clickable heat map: one cell per combination, plus your current model marked on it.
3. Read the sentence: it says how many points the best combination gains over yours. Below 0.005 it says so.
4. Click the best cell, then Appliquer cette combinaison.

1. Cross-validate the new model too. Then answer the only question that matters: is the gain from tuning bigger than the fold-to-fold noise you measured in Step 3?

Step 6: The Coach, and the model card10 min

Back on any results screen, the Coach block reads what is on screen and proposes at most six things. Its thresholds are the same as the "quality of fit" verdict just above it, so it never contradicts the screen.

Rule	Fires when	One-click remedy
overfitting	train / test gap above 0.10	max depth halved, or neighbours doubled, or C divided by 4
underfitting	training score too low	max depth + 5, or trees doubled
low score	test low, no gap, not underfitting	try a random forest
missing scaling	no standardisation on a distance-based algorithm	turn standardisation back on
imbalance / majority	smallest class under a third of the largest	rebalance the classes
leak	a leak warning was raised	drop the offending variable

Apply one. It creates a model named Coach: ..., comparable with the others. Then finish with the card:

1. Results screen, Fiche tab. Intended use, data, split actually applied, preparation, algorithm and parameters, metrics, and computed limits.
2. Click Désigner champion on your best model: one champion per project and per task, and the badge follows it into the comparison and the history.
3. Download the card in Markdown and paste its limits section into the Conclusion et limites section of your report. Then add the one limit the card cannot know: the truncated target.