lalala404/healformer-multiverse-maps

HealFormer Multiverse Maps

Targets-only HEALPix simulations used by HealFormer v0.2.0 for weak-lensing mass mapping. Each row contains one targets array in NESTED ordering with three float64 channels: [gamma1, gamma2, kappa].

Config	Nside	Train	Validation	Published mask/noise
nside256	256	9,000	1,000	none; generated at runtime
nside1024	1024	1,008	112	none; generated at runtime

The nside1024 subset is a deterministic capacity-limited 9:1 sample selected with seed 20260813. Masks and shape noise are intentionally not baked into the dataset. HealFormer's public processing pipeline generates them on the fly.

Stream without downloading the dataset

from datasets import load_dataset
import numpy as np

stream = load_dataset(
    "lalala404/healformer-multiverse-maps",
    "nside256",
    split="validation",
    streaming=True,
)
row = next(iter(stream))
targets = np.asarray(row["targets"], dtype=np.float64)
gamma1, gamma2, kappa_true = targets

assert targets.shape == (3, 12 * 256**2)

Use config nside1024 for the projection model. Streaming returns an IterableDataset; it does not download all Parquet shards.

Integrity and reproducibility

release-manifest.json records every published Parquet path, row count, byte size, and SHA-256 checksum. The release contains 657 Parquet shards totaling 293,419,758,199 bytes; the largest shard is 505,241,872 bytes.

The notebook in HealFormers v0.2.0 combines this dataset with the public model repositories to reproduce map residuals, hist2d distributions, power-spectrum ratios, and cross-correlation coefficients. Fixed evaluation masks and random seeds are declared by the notebook rather than hidden in the data.

License and citation

The dataset is released under CC0 1.0.

Yihe Wang and Yu Yu, Advancing weak lensing mass mapping with a mask-aware HEALPix transformer, arXiv:2603.25471.