This depends on the type of weight:

• Importance weights only affect the model estimation and supervised recipes steps. They are not used with yardstick functions for calculating measures of model performance.

• Frequency weights are used for all parts of the preprocessing, model fitting, and performance estimation operations.

(This is form the blog post but it should be better documented in hardhat).

So we should do this but just for frequency weights (which are less likely to be used with boosting.

@topepo @simonpcouch I've updated the PR to only pass freq weights to the internal validation set. This is done by delaying the conversion of the weights to numeric/integer till pass to as_xgb_data()

library(parsnip)

freq_weights <- hardhat::frequency_weights(1:32)

mtcar_x <- mtcars[, -1]
mtcar_mat <- as.matrix(mtcar_x)

set.seed(1)
val_freq_wts <- parsnip:::as_xgb_data(mtcar_mat, mtcars$mpg, weights = freq_weights, validation = 1/10)
xgboost::getinfo(val_freq_wts$watchlist$validation, "weight")
#> [1]  3 17 26


imp_wts <- hardhat::importance_weights(1:32)

mtcar_x <- mtcars[, -1]
mtcar_mat <- as.matrix(mtcar_x)

set.seed(1)
val_freq_wts <- parsnip:::as_xgb_data(mtcar_mat, mtcars$mpg, weights = imp_wts, validation = 1/10)
xgboost::getinfo(val_freq_wts$watchlist$validation, "weight")
#> NULL

^{Created on 2022-09-01 by the reprex package (v2.0.1)}