cran-comments.md

Test environments

• local Apple Silicon (M1), macOS 26.5.1 Tahoe, R 4.6.0
• GitHub Actions CI:
  • ubuntu-24.04 (release, devel)
  • windows-latest (release)
  • macOS-latest (release)
• R-hub
  • linux (R-devel)
  • macos (R-devel)
  • windows (R-devel)

R CMD check results

0 errors | 0 warnings | 0 notes

Minimum R version bumped from 3.5 to 4.1.0 due to use of the \() lambda shorthand syntax introduced in R 4.1.0.

Comments

This is a maintenance release that fixes several bugs in train_spectra(), test_spectra(), predict_spectra(), and internal utility functions:

• The final RF model was trained with ntree = tune.length (≤5 trees) instead of the standard 500; the tuned mtry was also not applied to the final model
• The final SVM model incorrectly included unique.id as a predictor
• The final PLS model used ncomp = tune.length instead of the best value identified during training iterations; predict_spectra() now reads ncomp directly from the model object
• train_spectra() crashed with undefined columns selected when cv.scheme was used, due to column indices computed before format_cv() removed the genotype column
• When cv.scheme is used, the final model was trained on incorrect data (either all trials combined or trial1 alone); it now calls format_cv() to use the scheme-appropriate training set
• set.seed() was called after createDataPartition(), making stratified train/test splits non-reproducible even when a seed was set
• best.model.metric and seed were accepted by test_spectra() but never forwarded to train_spectra(), silently ignoring user input
• SVM importance output was corrupted (1×1 matrix instead of NULL) when running with multiple pretreatments
• Invalid ntree and mtry arguments were passed to predict.randomForest(), where they are silently ignored

All existing user-facing APIs remain fully compatible.

Downstream dependencies

There are currently no downstream dependencies for this package.