48 Kernel models in plain terms

What people usually ask

Try an SVM on my data.

What comes back

A classifier fitted to a continuous viscosity column with no scaling, and either an error or an accuracy that means nothing.

Ask this instead

I have visc.csv: 30 runs, columns conc_wt_pct, temp_C and viscosity_mPas. The target is continuous, so this is regression, not classification.
1. A Pipeline of StandardScaler then SVR. Choose C from [0.1, 1, 10, 100] and epsilon from [0.01, 0.1, 1] with GridSearchCV, 5 folds, scoring neg_root_mean_squared_error. Print the chosen values and the cross validated RMSE.
2. A GaussianProcessRegressor with kernel ConstantKernel() * RBF() + WhiteKernel(), inputs scaled the same way, normalize_y=True. Print the fitted kernel with its length scale and noise level.
3. For the Gaussian process, plot the mean prediction against conc_wt_pct at temp_C = 25 with a band of two standard deviations, the measured points on top.
4. Under the plot, state that the band is what the chosen kernel implies, not a measured error: it narrows near points and widens between them.
Report both RMSEs beside a mean predictor under the same folds. Do not drop rows. Do not plot or report any value outside the measured range of either input.

Attach: The run table as CSV.

Why it works: Scaling sits inside the pipeline so the folds stay honest, and the band is named for what it is before anyone reads it as an error bar.