Palmer penguins

344 penguins of three species measured at Palmer Station, Antarctica, 2007–2009: bill, flipper, body mass, sex, island. Small and clean; the classification data for the week-3 seminar.

Provenance and license	Gorman, Williams and Fraser (2014), Palmer Station Long Term Ecological Research program; packaged by Horst, Hill and Gorman as palmerpenguins. The R package and its data are CC0.
R access	palmerpenguins::penguins. Use the qualified name: modeldata also exports a penguins (7 columns, no year) and masks it when attached later.
Unit of observation	A penguin (344 rows), 8 columns: species, island, bill_length_mm, bill_depth_mm, flipper_length_mm, body_mass_g, sex, year.
Target and loss	Week 3: species, Adelie against Chinstrap (219 birds after dropping incomplete measurements), 0–1 loss or a cost-weighted loss with the costs stated.
Available at prediction time	The measurements and island, sex, year. In this dataset island completely separates Chinstrap from Gentoo (Chinstraps are on Dream, Gentoos on Biscoe; Adelies occur on all three islands), so a model with island in it is solving a different, easier task; say so if a student uses it.
Permitted predictors	Week 3: bill_length_mm and bill_depth_mm (the default), flipper_length_mm alone (the weak model for the ROC curve).
Missingness	Two birds lack all four measurements; eleven lack sex. drop_na() on the columns used.
Split unit and seed	Birds. Week 3 fits on all 219 birds with no split: the point is the mechanism, and every number in the seminar is computed on the training rows (apparent performance; the validation week says what that flatters).
Baseline	Predict the majority class, Adelie: accuracy 0.69 on the two-species task.
Complication it carries	None social. Near-separability: the bill model has six errors in 219 birds and glm warns about fitted probabilities of 0 or 1, a live example of the lecture’s “too easy” slide.
Limitations	Three islands, three seasons, one research program: a clean dataset with no deployment story, which is why it is used when the mechanism is the lesson.
Fictional?	No.

Loader

library(palmerpenguins)
pg <- palmerpenguins::penguins |>
  dplyr::filter(species != "Gentoo") |>
  tidyr::drop_na(bill_length_mm, bill_depth_mm, flipper_length_mm) |>
  dplyr::mutate(chinstrap = as.numeric(species == "Chinstrap"))