generatedata

Datasets produced by the generatedata library. Every dataset is a start/target pair: start holds perturbed or noisy points, target holds the corresponding clean points on the manifold of interest. The pair defines a map — denoising, projection, or the flow of a dynamical system — and the library exists to make that map the same shape across very different data sources.

Loading

from generatedata import load_data

load_data.data_names()               # what is in this snapshot
data = load_data.load_data('MNIST')  # {'info': ..., 'start': df, 'target': df}
X, Y = load_data.load_data_as_xy('MNIST')

Nothing here requires a HuggingFace account or token.

Layout

The repository is flat, and info.json is the index — a dict from dataset name to that dataset's metadata (shapes, generating parameters, seeds, provenance). Reading it is how the library discovers what exists, so an unreferenced file is invisible to load_data.

File	Contents
info.json	index: dataset name → metadata
<name>_start.parquet	start points, one row per sample
<name>_target.parquet	target points, one row per sample
<name>_info.json	that dataset's metadata on its own
<name>_weights.npy	whest only: the flattened MLP weights
<name>_final_means.npy	whest only: the labels
<name>_ut_fixed.npy	whest only: the cheap baseline estimate
<name>_all_layer_means.npy	whest only: per-layer mean activations

The whest_* datasets are stored as .npy rather than parquet because a single row is an entire weight tensor — hundreds of thousands of floats — which is not a sensible DataFrame. The library memory-maps them, so a 90 MB array costs no RAM until rows are touched.

Versioning

Each upload is a commit, tagged vYYYYMMDD_HHMMSS. The library pins a commit hash in generatedata/config.py, so a given release of the library always reads exactly the same bytes. Older snapshots stay reachable by their hash or tag.

Provenance and licensing

The library code is BSD-3-Clause. Individual datasets carry their own upstream terms, recorded per dataset in info.json:

• MNIST, EMNIST, KMNIST, FashionMNIST, CIFAR-10 — standard research datasets, reprocessed into start/target form.
• MNIST-1D — Greydanus.
• Long Range Arena — the lra_* datasets are native reimplementations of the benchmark tasks.
• whest_* — the ARC White-Box Estimation Challenge. Rows sourced from aicrowd/arc-whestbench-public-2026 are marked source: official:* in info.json, with the pinned upstream revision and license: cc-by-4.0; the rest are generated from scratch.