lalala404/healformer-multiverse-maps

HealFormer Multiverse Maps

This repository provides the public spherical weak-lensing simulations, fixed survey masks, representative samples, and 100-sample diagnostic arrays used by HealFormer (Mask-Aware HEALPix Transformer).

Source code · Nside=256 model · Nside=1024 model · Executable notebook · Paper · arXiv

What is included

Each dataset row contains one targets array in HEALPix NESTED ordering with three float64 channels:

targets[0] = gamma1
targets[1] = gamma2
targets[2] = kappa
shape      = (3, 12 * nside**2)

The Parquet rows contain physical full-sky maps only. Survey masks and shape noise are deliberately not baked into the training rows: the HealFormers pipeline applies a selected mask and deterministic noise recipe at runtime. This keeps one source map reusable across controlled mask/noise experiments.

Config	Nside	Pixels per map	Train	Validation	Published evaluation masks
nside256	256	786,432	9,000	1,000	Fixed KiDS, DES, DECaLS, Planck
nside1024	1024	12,582,912	1,008	112	Fixed DECaLS only

The nside1024 rows are a deterministic, capacity-limited 9:1 subset selected with seed 20260813.

Stream a sample

Install the Hugging Face Datasets dependency:

python -m pip install "datasets>=2.19,<4" numpy

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

print(targets.shape)  # (3, 786432)

Use config nside1024 with the projection checkpoint. Streaming returns an IterableDataset and reads shards as needed; it does not download the full dataset first.

For release-pinned access with schema and integrity validation, install healformers>=0.2,<0.3 and use:

from healformers import get_public_release

sample = next(iter(get_public_release(256).stream_validation()))
print(sample.sample_id, sample.gamma1.shape, sample.kappa.shape)

Masks, noise, and evaluation artifacts

The evaluation/ directory contains:

• fixed-survey-masks.npz and its provenance manifest;
• one representative inference sample for each resolution; and
• raw per-sample HealFormer and spherical Kaiser--Squires power-ratio and cross-correlation arrays over 100 fixed-mask validation skies.

The published evaluation recipe uses NESTED ordering, shape-noise sigma 0.4, galaxy density 30 arcmin^-2, shared two-channel noise, sampling seed 20260813, and no map rotation. The source targets remain unchanged.

At Nside=256, one model checkpoint is evaluated on fixed KiDS, DES, DECaLS, and Planck footprints. At Nside=1024, evaluation uses only the fixed DECaLS-labeled footprint and the projection checkpoint.

DECaLS label note: the footprint labeled “DECaLS” in the paper and released artifacts is the combined DECaLS+DES footprint. This is a label-recording offset; the stored mask, calculations, method, and conclusions are unchanged.

Reproduce the figures

The HealFormers quickstart notebook loads the public checkpoints and this repository to reproduce:

• input shear, mask, convergence, and normalized residual maps;
• visible-pixel hist2d comparisons;
• 100-sample power-spectrum ratios; and
• 100-sample harmonic cross-correlation coefficients with one-sigma bands.

By default the notebook renders the checksummed 100-sample arrays. Set HEALFORMER_RECOMPUTE_ENSEMBLE=1 to sample 100 validation maps with seed 20260813, stream their targets, apply the released mask/noise recipe, and independently rerun HealFormer and spherical Kaiser--Squires.

Integrity

release-manifest.json records every Parquet path, row count, byte size, and SHA-256 digest. The release contains 657 Parquet shards totaling 293,419,758,199 bytes; the largest shard is 505,241,872 bytes. Evaluation artifacts and mask source files are independently checksummed in their own manifests.

Intended use and limitations

This dataset is intended for research on simulated spherical weak-lensing mass mapping, model training, controlled mask/noise studies, and reproduction of the released diagnostics. It is not a catalog of real survey observations and does not encode survey-specific photometric-redshift, shear-calibration, selection, or spatially varying noise systematics. Validate any derived model before use on observational data.

License and citation

The current dataset release is distributed under the Apache License 2.0. See LICENSE and NOTICE for the complete terms and attribution. Earlier immutable revisions published under CC0 retain their original terms.

@article{wang2026advancing,
  title={Advancing weak lensing mass mapping with a mask-aware HEALPix transformer},
  author={Wang, Yihe and Yu, Yu},
  journal={Physical Review D},
  volume={113},
  number={4},
  pages={043553},
  year={2026},
  publisher={APS},
  doi={10.1103/kc9z-jllp}
}