MLP Probe (1-hidden-layer non-linear classifier/regressor)

Fit a one-hidden-layer MLP on top of a frozen embedding to predict a target. Supports regression (mode=regress, R²) and classification (mode=classify, accuracy). 1000-resample bootstrap CI on the held-out metric. Stratified 80/20 train/test split (by class for classify, by quintile of value for regress). sklearn MLPRegressor/MLPClassifier backend.

ID	mlp_probe
Domain	ml
Category	ml
Type	tabular-to-metric
Version	0.1.0

view yaml ◦view github ◦

Parameters

mode: string · default regress · "regress" → R² on a continuous target. "classify" → accuracy on an integer-labeled target.
hidden_units: integer · default 64 · Width of the single hidden layer (ticket-locked at 64).
dropout: float · default 0.1 · sklearn does not implement dropout; mapped to L2 penalty `alpha = dropout * 1e-3`. Tracked separately so a future PyTorch backend can implement real dropout without changing recipes.
train_frac: float · default 0.8 · Stratified train-set fraction.
seed: integer · default 42 · Random seed (split + MLP init + bootstrap).
n_bootstrap: integer · default 1000 · Number of bootstrap resamples for the CI.
max_iter: integer · default 200 · Max solver iterations.
solver: string · default adam · "adam" for 10k+ row datasets (default); "lbfgs" for small synthetic targets where adam under-converges (unit tests).