Allocation stratified sampling with R

This is my dataset

set.seed(2026)

n <- 1000

img_id <- sprintf("IMG_%04d", seq_len(n))

strata <- sample(c("A", "B", "C"), size = n, replace = TRUE, prob = c(0.8, 0.1, 0.1))

split <- sample(c("train", "test", "val"), size = n, replace = TRUE, prob = c(0.7, 0.2, 0.1))

dataset <- data.frame(img_id, split, strata)

dataset %>% count(split, strata)

  split strata   n
1  test      A 154
2  test      B  21
3  test      C   9
4 train      A 573
5 train      B  64
6 train      C  77
7   val      A  79
8   val      B  12
9   val      C  11

It's an unbalanced dataset. I am trying to proportionate the stratification, as it is proposed here. I want to sample 300 images from training, 80 from val and 80 from test.

allocation <- dataset %>%
  count(split, strata, name = "Nh") %>%
  group_by(split) %>%
  mutate(N_split = sum(Nh)) %>%
  ungroup() %>%
  left_join(target, by = "split") %>%
  mutate(
    nh = round(Nh / N_split * n_target)
  ) %>%
  arrange(split, strata)

  split strata    Nh N_split n_target    nh
  <chr> <chr>  <int>   <int>    <dbl> <dbl>
1 test  A        154     184       80    67
2 test  B         21     184       80     9
3 test  C          9     184       80     4
4 train A        573     714      300   241
5 train B         64     714      300    27
6 train C         77     714      300    32
7 val   A         79     102       80    62
8 val   B         12     102       80     9
9 val   C         11     102       80     9

Nh: available images per Strat, per split. N_split: sum of images per split. n_target: how many images I would like to have per split. nh: proportionated stratification.

Now, let's see the test split. For Strat B, there are still 12 images available (Nh - nh). I would like to make a condition for all strats: if less than 20 images were computed in nh, take all the available images and discount them from the strata with the biggest representation, in this case, strat A, in order to match with the desired amount of sampled images per split.