Top-p parameter for dials

top_p() creates a dials parameter for nucleus sampling (the share of probability mass the model samples from), for tuning llm_predict() specifications or building grids for optimize_grid(). Like temperature(), it applies to gpt-6-luna only when its reasoning effort is "none".

Usage

top_p(range = c(0, 1), trans = NULL)

Arguments

range

Range of values (default c(0, 1)).

trans

Optional transformation from the scales package; NULL (the default) means none.

Value

A dials quantitative parameter.

See also

Other integrations: as_dsprrr_metric(), as_dsprrr_traces(), as_ellmer_tool(), as_vitals_cost(), as_vitals_samples(), as_vitals_solver(), as_vitals_task(), create_search_tool(), llm_predict(), ragnar_tool(), reasoning_effort(), register_dsprrr_engine(), summarize_traces_df(), temperature(), use_dsprrr_template(), validate_workflow(), vitals_metrics

Examples

top_p()
#> Top P (quantitative)
#> Range: [0, 1]
top_p(range = c(0.5, 1))
#> Top P (quantitative)
#> Range: [0.5, 1]