Waterbirds (OCCAM layout)

This repository hosts the Waterbirds image files used in the OCCAM codebase (arXiv), laid out for experiments on subpopulation / group shifts, foreground-only, and background-only evaluation.

Original data and credit

The images come from the Waterbirds benchmark introduced with group distributionally robust optimization in:

Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, Percy Liang, Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization, arXiv:1911.08731.

Please cite that work when using the original benchmark. Licensing and redistribution terms of the underlying images follow the original dataset / WILDS release; refer to the paper and official sources for details.

Folder layout (twelve subscenarios)

On the Hugging Face Files tab you should see twelve top-level folders (three per historical group_0 … group_3): the original scene, *_fg_only (foreground crop), and *_bg_only (background crop). Each triplet shares the same spurious-cue group:

Folder	Description
landbird_on_land	Original image (foreground + background); same spurious-cue group as historical group_0
landbird_on_water	Original image (foreground + background); same spurious-cue group as historical group_1
waterbird_on_land	Original image (foreground + background); same spurious-cue group as historical group_2
waterbird_on_water	Original image (foreground + background); same spurious-cue group as historical group_3
landbird_on_land_fg_only	Foreground-only crop for the same group as landbird_on_land
landbird_on_water_fg_only	Foreground-only crop for the same group as landbird_on_water
waterbird_on_land_fg_only	Foreground-only crop for the same group as waterbird_on_land
waterbird_on_water_fg_only	Foreground-only crop for the same group as waterbird_on_water
landbird_on_land_bg_only	Background-only crop for the same group as landbird_on_land
landbird_on_water_bg_only	Background-only crop for the same group as landbird_on_water
waterbird_on_land_bg_only	Background-only crop for the same group as waterbird_on_land
waterbird_on_water_bg_only	Background-only crop for the same group as waterbird_on_water

Background-only crops are paired with fg+bg composites via metadata.csv at the dataset root (img_filename ↔ place_filename; same basename under *_bg_only as under the matching fg+bg subscenario). The upload script copies pixels from Places 256 (data/dataset/places/data_256_standard by default, configurable via --places-root) using place_filename (see sync_bg_only_subscenarios_from_metadata in occam/datasets/waterbirds_metadata.py).

Class labels inside 0/ and 1/

Each subscenario folder contains subfolders 0 and 1, which are the binary coarse bird-type labels used by OCCAM configs and ImageFolder-style loaders:

• 1 → landbird
• 0 → waterbird

(These are not the 200 fine-grained species names; they are the two high-level types for the Waterbirds classification head in this benchmark.)

Foreground-only crops follow the deep feature reweighting setting; extraction follows Kirichenko, Izmailov & Wilson, Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations (arXiv:2204.02937; code).

Background-only crops use the same grouping as the original Waterbirds benchmark; they are distributed alongside the other subscenarios for analysis (e.g. background shift without the bird).

Hub Dataset Viewer (Subset = subscenario)

The dataset card YAML declares configs with one entry per subscenario. Each config sets data_dir to that folder so the Hub uses the built-in ImageFolder loader: one train split per subset, columns image and label (folder names 0 / 1; see above for bird-type meaning).

Example:

from datasets import load_dataset

ds = load_dataset("YOUR_ORG/waterbirds", "landbird_on_land_fg_only", split="train")

No trust_remote_code is required (datasets 4.x does not load Hub Python dataset scripts).

If the viewer still shows a single default subset after updating the card, delete any stale auto-generated data/ folder on the Hub Files tab (leftover from an older layout) and refresh the page.

metadata.csv

The repository includes metadata.csv at the root (WILDS-style columns: img_id, img_filename, y, split, place, place_filename). Use it to recover the original bird image path and background Places path for each composite. Under each Hub subscenario, image files are named from img_filename; the matching *_bg_only file uses the same basename so fg+bg and bg-only subsets stay aligned.

OCCAM codebase

Download scripts, configs, and full experiment documentation live in the OCCAM repo:

• https://github.com/AlexanderRubinstein/OCCAM/blob/main/README.md

The canonical download path in the codebase is scripts/download_datasets_and_checkpoints.py, which fetches this dataset from the Hub after installing UrbanCars / CounterAnimals from the shared Google Drive archive.

Citation (OCCAM)

If you use this exact packaging together with OCCAM, please also cite the OCCAM paper (HF paper page, arXiv).