Feature request: add supplementary data into train/test after resampling

I’d like to propose adding a utility function to {rsample} that allows appending new data to either the training or testing set of an rsplit or rset object after resampling.

This comes up in my workflow, for instance, to assess the marginal impact of a cohort on model performances (by training on data with and without the additional set, then comparing performance on a common test set).

I wrote a small helper function that could serve as a starting point, but it could likely be improved by leveraging the full internal structure and capabilities of {rsample}.

library(rsample)

add_data_rsample <- function(x, new_data, into = c("train", "test")) {
  stopifnot(is(x, "rsplit") | is(x, "rset"))

  into <- match.arg(into)

  if (is(x, "rsplit")) {
    if (into == "train") {
      y <- make_splits(dplyr::bind_rows(training(x), new_data), testing(x))
    }
    if (into == "test") {
      y <- make_splits(training(x), dplyr::bind_rows(testing(x), new_data))
    }
  }

  if (is(x, "rset")) {
    new_splits <- purrr::map(x$splits, ~ add_data_rsample(., new_data, into))
    y <- manual_rset(new_splits, x$id)
  }

  y
}

mt_cv <- vfold_cv(mtcars[1:24, ], v = 4)
mt_cv
#> #  4-fold cross-validation 
#> # A tibble: 4 × 2
#>   splits         id   
#>   <list>         <chr>
#> 1 <split [18/6]> Fold1
#> 2 <split [18/6]> Fold2
#> 3 <split [18/6]> Fold3
#> 4 <split [18/6]> Fold4
add_data_rsample(mt_cv, mtcars[25:32, ], into = "train") # + 8 rows in train
#> # Manual resampling 
#> # A tibble: 4 × 2
#>   splits         id   
#>   <list>         <chr>
#> 1 <split [26/6]> Fold1
#> 2 <split [26/6]> Fold2
#> 3 <split [26/6]> Fold3
#> 4 <split [26/6]> Fold4
add_data_rsample(mt_cv$splits[[1]], mtcars[25:32, ], into = "test") # + 8 rows in train
#> <Analysis/Assess/Total>
#> <18/14/32>

^{Created on 2025-07-10 with reprex v2.1.1}

I hope this function could be useful for others.

Thanks for your consideration, and thanks again for your great work on {rsample}!

Start with the rsplit and rset entry points in the example, then read how make_splits() and manual_rset() represent training and testing data. Compare the proposed helper’s behavior for both object types and both into values. Done means a supported utility preserves the resampling structure while appending rows to the selected set, with coverage for the resulting split sizes.