Add Parsnip Model

Add Parsnip Model

Create entirely new model specifications for the parsnip package. This skill covers creating new model types (like linear_reg(), boost_tree()) with their constructors, registration, and engine implementations.

Use this skill when: Creating a fundamentally new model type for parsnip.

For adding engines to existing models: See add-parsnip-engine skill instead.

Prerequisites

Before creating a new parsnip model, ensure you have:

R Package Development:

• Extension Prerequisites - Required for extension packages

• Development Workflow - Fast iteration practices

Parsnip Architecture:

• Model Specification System - Core parsnip concepts

• Fit and Predict Methods - Implementation fundamentals

Creating a New Model

Assess Model Complexity First

Before diving into implementation, determine the complexity of your model:

Simple Model:

• Single mode (regression OR classification, not both)

• 1-2 engines

• Few main arguments (1-3)

• Standard prediction types only

→ Follow streamlined approach: Focus on getting the basics right, avoid over-engineering

Complex Model:

• Multiple modes (regression AND classification)

• 3+ engines

• Many main arguments

• Custom prediction types or special encoding

→ Reference detailed guides for multi-mode handling, encoding options, and advanced patterns

Target files regardless of complexity:

• Extension development: 2-3 files (constructor, tests, optional README); acceptable to reach 4-6 if needed

• Source development: 2-4 files (constructor, data file, tests, optional engine docs); acceptable to reach 5-8 if needed

1. Design the Model Specification

Start here: Model Constructor Design

Decide on:

• Model name and function (e.g., sparse_reg())

• Which modes to support (regression, classification, both?)

• Main arguments (standardized across engines)

• Default engine

This step defines the user-facing API.

2. Understand the Registration System

Review: Model Specification System

Learn how parsnip's registration system works:

• Model environment and storage

• Engine registration database

• Mode handling

• Argument translation

3. Implement the Registration Sequence

Follow: Registration Sequence

Complete registration in the correct order: 1. set_new_model() - Declare model exists 2. set_model_mode() - Declare supported modes 3. set_model_engine() - Register each engine 4. set_dependency() - Package requirements 5. set_model_arg() - Argument translation 6. set_fit() - Fitting method 7. set_encoding() - Data conversion (if needed) 8. set_pred() - Each prediction type

4. Design Main Arguments

Plan carefully: Argument Design

Create standardized arguments that:

• Work across multiple engines

• Map to engine-specific parameters

• Integrate with tune package

• Follow tidymodels conventions

5. Implement Fit and Predict

Core implementation: Fit and Predict Methods

For each engine:

• Choose interface type (formula, matrix, xy)

• Implement data conversion

• Register fit method

• Register each prediction type with proper column naming

6. Handle Prediction Types

Standardize output: Prediction Types

Implement appropriate prediction types for each mode:

• Regression: numeric, conf_int, pred_int

• Classification: class, prob

• Survival: time, survival, hazard, linear_pred

• Quantile: quantile

7. Configure Mode Handling

If multi-mode: Mode Handling

For models supporting multiple modes:

• Register each mode separately

• Set mode-specific defaults

• Implement mode-specific prediction types

• Handle mode validation

8. Handle Encoding Options

For matrix/xy interfaces: Encoding Options

Configure how formulas are converted:

• Choose interface type

• Set indicator coding (traditional vs one-hot)

• Handle intercept

• Configure factor encoding

Testing Your Model

Essential tests:

• Model constructor creates correct object

• Setting engine works

• Setting mode works (if multi-mode)

• Formula and xy interfaces equivalent

• Each prediction type returns correct format

• Factor handling works correctly

• Error messages are clear

Test each engine separately:

test_that("lm engine works", {
  skip_if_not_installed("lm_package")

  spec <- my_model() |> set_engine("lm")
  fit <- fit(spec, y ~ x, data = data)
  expect_s3_class(fit, "model_fit")

  preds <- predict(fit, data)
  expect_named(preds, ".pred")
})

Contributing to Parsnip Source

For PRs to tidymodels/parsnip:

Additional resources for source development:

• Best Practices (Source) - Parsnip-specific patterns

• Troubleshooting (Source) - Common issues

• Testing Patterns (Source) - Comprehensive tests

Key differences from extensions:

• Can use internal functions (:::)

• Follow parsnip file organization (R/[model].R, R/[model]_data.R)

• Add to parsnip documentation

• More comprehensive testing required

• Consider existing parsnip conventions

Example: Creating sparse_reg()

Hypothetical new model for sparse regression:

Note: This example shows extension development patterns. For source development, omit parsnip:: prefixes and use internal functions as shown in source-guide.md.

1. Constructor (R/sparse_reg.R):

   sparse_reg <- function(mode = "regression",
                          penalty = NULL,
                          sparsity_threshold = NULL,
                          engine = "glmnet") {
   
     args <- list(
       penalty = rlang::enquo(penalty),
       sparsity_threshold = rlang::enquo(sparsity_threshold)
     )
   
     parsnip::new_model_spec(
       "sparse_reg",
       args = args,
       eng_args = NULL,
       mode = mode,
       user_specified_mode = FALSE,
       method = NULL,
       engine = engine,
       user_specified_engine = !missing(engine)
     )
   }
   

2. Registration (in .onLoad() or package setup):

   # Declare model
   parsnip::set_new_model("sparse_reg")
   parsnip::set_model_mode("sparse_reg", "regression")
   
   # Register glmnet engine
   parsnip::set_model_engine("sparse_reg", "regression", "glmnet")
   parsnip::set_dependency("sparse_reg", "glmnet", "glmnet", "regression")
   
   # Translate arguments
   parsnip::set_model_arg(
     model = "sparse_reg",
     eng = "glmnet",
     parsnip = "penalty",
     original = "lambda",
     func = list(pkg = "dials", fun = "penalty"),
     has_submodel = TRUE
   )
   
   # Fit method
   parsnip::set_fit(
     model = "sparse_reg",
     eng = "glmnet",
     mode = "regression",
     value = list(
       interface = "matrix",
       protect = c("x", "y"),
       func = c(pkg = "glmnet", fun = "glmnet"),
       defaults = list(family = "gaussian")
     )
   )
   
   # Predictions
   parsnip::set_pred(
     model = "sparse_reg",
     eng = "glmnet",
     mode = "regression",
     type = "numeric",
     value = list(
       pre = NULL,
       post = NULL,
       func = c(fun = "predict"),
       args = list(
         object = rlang::expr(object$fit),
         newx = rlang::expr(as.matrix(new_data)),
         type = "response"
       )
     )
   )
   

3. Testing:

   test_that("sparse_reg constructor works", {
     spec <- sparse_reg()
     expect_s3_class(spec, "sparse_reg")
     expect_equal(spec$mode, "regression")
     expect_equal(spec$engine, "glmnet")
   })
   
   test_that("sparse_reg fits and predicts", {
     skip_if_not_installed("glmnet")
   
     spec <- sparse_reg(penalty = 0.1) |> set_engine("glmnet")
     fit <- fit(spec, mpg ~ ., data = mtcars)
   
     expect_s3_class(fit, "model_fit")
   
     preds <- predict(fit, mtcars[1:5, ])
     expect_s3_class(preds, "tbl_df")
     expect_named(preds, ".pred")
     expect_equal(nrow(preds), 5)
   })
   

Common Pitfalls

Prioritize correctness over structure. Focus on getting the implementation right before worrying about file organization.

Code correctness issues (fix first): 1. Incorrect column names - Follow .pred naming conventions strictly (.pred, .pred_class, .pred_lower) 2. Wrong interface type - Match engine's expected input format (formula vs matrix vs xy) 3. Inconsistent argument naming - Use tidymodels standards (penalty, mixture), not engine names 4. Missing prediction post-processing - Ensure output format matches parsnip conventions

Implementation completeness (fix second): 5. Incomplete registration - Must complete full sequence for each engine (set_new_model → set_model_mode → set_model_engine → set_dependency → set_model_arg → set_fit → set_pred) 6. Missing mode registration - Must register modes explicitly with set_model_mode() 7. No argument translation - Main arguments must map to engine arguments via set_model_arg() 8. Insufficient testing - Test all modes, engines, prediction types, and both fit() and fit_xy()

Structural concerns (address last): 9. Too many files - Keep to 2-3 files for extensions, 2-4 for source (see File Discipline section)

When to Create a New Model

INSTRUCTIONS FOR CLAUDE:

Before implementing, verify this is truly a NEW model type. If the user requests:

• A different computational engine for an existing model (e.g., "add xgboost to boost_tree") → Stop. Politely explain this should use the add-parsnip-engine skill instead

• A minor variation that could be an engine-specific argument → Stop. Suggest using engine-specific arguments rather than creating a new model

• Something that duplicates an existing parsnip model → Stop. Point them to the existing model

Only proceed with implementation if it's genuinely a new model type that doesn't exist in parsnip.

Create a new model when:

• The algorithm is fundamentally different from existing models

• It serves a distinct use case

• It has unique prediction types

• It fills a gap in parsnip's model coverage

Don't create a new model when:

• It's just a different engine for an existing model type → Use add-parsnip-engine instead

• It's a minor variation of an existing model → Consider engine-specific arguments instead

Examples:

• ✓ survival_reg() - New outcome type (censored data)

• ✓ naive_bayes() - Distinct algorithm family

• ✗ Random forest with different package → Add engine to rand_forest()

• ✗ Linear regression with different penalty → Add engine to linear_reg()

File Discipline

Keep implementations focused and avoid creating unnecessary files.

Target file counts:

Extension development:

• R/[model_name].R - Model constructor

• tests/testthat/test-[model_name].R - Tests

• README.md - Only if needed for package users

• Total: 2-3 files (acceptable to reach 4-6 if implementation requires it)

Source development:

• R/[model_name].R - Model constructor

• R/[model_name]_data.R - Engine registrations

• tests/testthat/test-[model_name].R - Tests

• man/rmd/[model_name]_[engine].Rmd - Engine docs (optional)

• Total: 2-4 files (acceptable to reach 5-8 if implementation requires it)

Do not create:

• Implementation notes or summaries (IMPLEMENTATION_NOTES.md, IMPLEMENTATION_SUMMARY.md)

• Usage example files (example_usage.R, examples.R)

• Separate documentation files (DOCUMENTATION.md, USAGE.md)

• Development guide files (DEVELOPMENT.md, GUIDE.md)

• Testing guide files (TESTING.md, TEST_GUIDE.md)

• Changelog files (CHANGELOG.md, NEWS.md unless source development)

• Configuration files (CONFIG.md, SETUP.md)

• Workflow files (WORKFLOW.md, PROCESS.md)

• Debug or log files (DEBUG.md, LOG.md)

• Status or progress files (STATUS.md, PROGRESS.md)

• TODO or task files (TODO.md, TASKS.md)

• Multiple README variants (README_DEV.md, README_TECHNICAL.md)

Instead:

• Put usage examples in roxygen @examples tags

• Put implementation notes in roxygen @details tags

• Put development notes in comments within code

• Document design decisions in commit messages

• Use vignettes for comprehensive usage guides (if creating a package)

Related Skills

• add-parsnip-engine - Add computational engines to your new model

• add-dials-parameter - Define tunable parameters for model arguments

• add-recipe-step - Preprocess data before model fitting

• add-yardstick-metric - Evaluate model predictions with custom metrics

Next Steps

After creating your model:

1. Test thoroughly - Both extension and source tests
2. Document - Add examples and usage guidance
3. Share - Consider contributing to parsnip
4. Maintain - Keep up with engine package updates

For questions or contributions, see:

• Tidymodels GitHub

• Tidymodels Community