LoginSign Up
← Back to all sparks
D

dials

ANALYTICS
Velocity0.0

Create and manage tuning parameter grids and ranges for model tuning in R

github.com/tidymodels/dials ↗

dials is quietly registering the tuning parameters for tidymodels' deep-learning push

tidymodelshyperparametersdeep-learninggrid-searchr-stats
◆Current state
dials defines the parameter objects and grid constructors that tidymodels tunes over, which makes its release notes a reliable early read on what the rest of the stack is about to support. The last two releases are dominated by attention-model parameters — SAINT and tabular deep learning via brulee, TabPFN via parsnip's tab_pfn() — alongside catboost parameters for bonsai and calibration parameters for tailor.
◆Where it's heading
The grid machinery itself is settled: grid_space_filling() consolidated the older designs, and the grid_*() functions now error rather than warn on the wrong size argument. What keeps moving is the parameter catalog, and it is moving toward neural and foundation-model territory that tidymodels historically left alone. Error-message quality is a steady secondary theme.
◆Prediction
Expect further parameter objects to land ahead of the parsnip and brulee releases that use them — the attention and tabular-foundation-model work in flight is the clearest thing the entries point to.

◆Recent moves

  1. 1mo ago

    Attention and tabular deep-learning parameters for brulee models

    A dozen parameters for SAINT and the tabular package's deep-learning models, plus catboost tuning objects for bonsai and survival support in splitting_rule(). Registration work that lands before the models it serves are usable.

    View source ↗
  2. 4mo ago

    Parameters for ordinal_reg() and the tab_pfn() foundation model

    Adds the tuning surface for parsnip's new ordinal regression and for tab_pfn(), including softmax temperature and training-set limits — a tabular foundation model getting first-class tidymodels parameters. Error messages across the package also get more specific.

    View source ↗
  3. 11mo ago

    prop_terms() for supervised feature selection recipes

    A new parameter for supervised feature selection, plus inclusive endpoints on the limit parameters and a real default range for batch_size() in place of an unknown one.

    View source ↗
  4. 1y ago

    Calibration method parameters for classification and regression

    Adds cal_method_class() and cal_method_reg() so calibration choice becomes tunable rather than fixed — the postprocessing theme that started with the tailor parameters in 1.4.0.

    View source ↗
  5. 1y ago

    Postprocessing parameters added; grid size mismatches now error

    Brings tailor's equivocal-zone and numeric-range parameters into the tunable set, translates all messaging to cli, and turns wrong grid-size arguments from warnings into errors. Small-design space-filling behavior also improves.

    View source ↗
  6. 2y ago

    grid_space_filling() consolidates the space-filling designs

    Replaces grid_max_entropy() and grid_latin_hypercube() with a single entry point covering several precomputed design families. The consolidation that makes the grid API stable enough to stop changing.

    View source ↗