Build a tidymodels parameter set for a module

module_parameters() returns a dials::parameters() set describing values of a module that can be tuned. Pass it to optimize_grid() as parameters to search a regular or random grid over those values.

Usage

module_parameters(
  module,
  model = NULL,
  include = NULL,
  exclude = c("id", "instructions", "instructions_suffix")
)

Arguments

module

A module, such as one created with module().

model

Optional model name, such as "gpt-6-luna". For a reasoning model, temperature and top_p are replaced by reasoning_effort.

include

Optional character vector of parameter names to keep.

exclude

Character vector of parameter names to drop when include is NULL. The default drops id, instructions and instructions_suffix.

Value

A dials::parameters() object, empty when nothing tunable is found.

Details

Candidate parameters come from:

• single values in module$config and module$config$params;

• the parameters of trials recorded by optimize_grid();

• enum inputs of the signature, as input_<name> with the enum levels;

• runtime settings with default ranges: temperature and top_p in [0, 1], frequency_penalty and presence_penalty in [-2, 2], and max_output_tokens in [32, 4096].

Numeric values become quantitative parameters spanning the observed range (plus or minus 0.1 around a single value); character and logical values become qualitative parameters. For a reasoning model (see is_reasoning_model()), temperature and top_p are dropped and reasoning_effort ("low", "medium", "high") is added.

Only runtime settings, instructions, instructions_suffix and template change what optimize_grid() sends to the model. Other parameters, such as input_<name> or internal config fields like .module_kind, are stored in the module's config without effect, so use include to keep the ones you mean to tune.

See also

Other grid search: GridSearchTeleprompter(), module_metrics(), module_trials(), optimize_grid()

Examples

mod <- module(
  signature("text -> sentiment"),
  config = list(temperature = 0.2)
)
module_parameters(mod, include = c("temperature", "top_p"))
#> Collection of 2 parameters for tuning
#> 
#>   identifier        type    object
#>  temperature temperature nparam[+]
#>        top_p       top_p dparam[+]
#> 

# Reasoning models tune reasoning_effort instead of temperature
module_parameters(
  mod,
  model = "gpt-6-luna",
  include = c("temperature", "reasoning_effort")
)
#> Collection of 1 parameters for tuning
#> 
#>        identifier             type    object
#>  reasoning_effort reasoning_effort dparam[+]
#>