aryana-haghjoo/romansr-data

specsr-roman training data — Roman grism spectra from OpenUniverse2024

Paired low-resolution Roman grism spectra and their noiseless ground-truth SEDs, for training and evaluating spectral super-resolution. Built with specsr-roman; the models trained on it are at aryana-haghjoo/roman-spectral-superresolution.

Spectra	36,404
Unique galaxies	15,434
Redshift	0.036 – 3.005 (median 0.765; 35 % above z = 1)
Depth	AB(H158) 14.7 – 22.5
Pointings	5 visits × 18 SCAs, healpix 10307
Median extraction S/N	0.50

Files

File	Size	What
ou2024_h10307_dataset.npz	271 MB	the dataset
tutorial/ou2024_h10307_tutorial.npz	3.8 MB	a 512-row held-out sample, for the tutorial notebook
splits/group_split_41dd131a0fd7750d29dd54c4920cb871.npz	0.3 MB	the canonical train/test split
pred_cache.npz	335 MB	frozen test-split predictions of the published chain

pred_cache.npz lets you reproduce every published figure and metric without a GPU or the model weights:

import numpy as np
from huggingface_hub import hf_hub_download
from specsr_roman.evaluation import line_amplitude_recovery, redshift_summary

c = np.load(hf_hub_download("aryana-haghjoo/romansr-data", "pred_cache.npz",
                            repo_type="dataset"))
redshift_summary(c["z_pred"], c["z_true"])
# {'nmad': 0.0065, 'median_abs_dz': 0.0047, 'catastrophic_frac': 0.0509, 'n': 7334}

The tutorial subset

If you want to try the model rather than train one, start with the 3.8 MB sample instead of the full 271 MB:

path = hf_hub_download("aryana-haghjoo/romansr-data",
                       "tutorial/ou2024_h10307_tutorial.npz", repo_type="dataset")

512 spectra of 476 galaxies, same schema as the full file plus a source_row column recording where each row came from in it.

Two properties of the sampling matter more than its size. It is drawn from the held-out side of the canonical object-id split, so metrics computed on it are honest out-of-sample numbers rather than a measure of memorisation. And it is a uniform random draw within that split — not a selection of bright objects or strong lines — so it keeps the population's real mix of recoverable and undetectable lines. That mix is exactly what the recoverability-binned line-recovery metric needs in order to mean anything: a hand-picked set of pretty spectra would make any model look good.

It drives the getting-started notebook, which goes from install to the published numbers in about two minutes on a CPU. Rebuild the subset with python scripts/make_tutorial_dataset.py.

Schema

Every row lives on two shared wavelength grids, so any two spectra are directly comparable.

Key	Shape	Meaning
flux_low	(36404, 864)	extracted grism spectrum, native ~10.764 Å sampling
flux_low_err	(36404, 864)	its 1σ uncertainty
flux_high	(36404, 2500)	ground-truth SED — the super-resolution target
redshift	(36404,)	true redshift
ids	(36404,)	OU2024 object_id — split on this
phot	(36404, 14)	catalogue fluxes (see band order below)
ab_h158	(36404,)	AB magnitude in H158
snr	(36404,)	median extraction S/N
visit, sca	(36404,)	provenance
wavelength_low	(864,)	10000–19290 Å
wavelength_high	(2500,)	10000–19300 Å
phot_bands	(14,)	photometry column names

There is no flux_high_err: the targets are noiseless simulated SEDs.

Flux units are internal to the simulation and are not calibrated across flux_low and flux_high — they differ by roughly 20 orders of magnitude. Only the shape is meaningful; normalise each spectrum before use, which is what specsr_roman.data.RomanFixedGridDataset does.

Photometry band order

0-5    LSST   u  g  r  i  z  y
6-13   Roman  R062  Z087  Y106  J129  W146  H158  F184  K213

⚠️ Two traps.

1. OU2024 uses the old Roman band names, which collide with the current WFI scheme: OU2024 R062 is 0.62 µm (current F062) and OU2024 W146 is the current wide filter R062. Map by central wavelength, never by name.
2. Feeding a model every band means giving it an effectively complete, noiseless SED — the redshift can be read off without the spectrum contributing anything, which measures the catalogue rather than the instrument. Restrict to bands that ship with the grism (Roman Medium tier: indices 8, 9, 11) and add realistic noise.

Splitting

Split by object, never by row. The same galaxy appears in several visits as independent noise realisations; a row-wise split puts one realisation in train and another in test, and measures memorisation.

from specsr_roman.data import get_or_make_group_split
train_idx, test_idx, _ = get_or_make_group_split(path, data["ids"])

Membership is a pure hash of the object id, so it is reproducible without the split file and stable under dataset growth. The included split file is the canonical one: 29,070 train / 7,334 test rows (12,336 / 3,098 galaxies).

How it was built

OpenUniverse2024 provides Roman direct images, truth catalogues and Diffsky SEDs, but no grism images. specsr-roman disperses the scene itself with grizli using the Wang et al. (2022) Roman.G150.conf, then extracts it back.

Because the same configuration disperses and extracts, input and target are self-consistent by construction: any residual is noise, blending or the extraction, not a mismatched instrument model.

Per (visit, SCA):

1. OU2024 H158 image → grizli direct frame (counts → e⁻/s, flat sky removed) plus an empty grism shell sharing its WCS.
2. photutils detection, KD-tree matched to the truth index.
3. Every source brighter than AB 23.0 dispersed with its true Diffsky SED.
4. Roman grism noise: Poisson (source + 0.57 e⁻/s/pix zodiacal) plus 8.5 e⁻ read noise, at a 301 s single-exposure depth.
5. Per target (AB < 22.5): re-dispersed alone and subtracted from the scene for exact contamination, then optimally extracted and flux-calibrated against a flat-f_λ pass through the same beam.

Reproduce with specsr-roman extract run then specsr-roman extract merge.

Limitations

• One simulation's manifold. Targets are Diffsky model SEDs with that simulation's line physics and dust treatment. A model can score well by learning the manifold rather than by measuring anything — see the prior-dominance audit in specsr-roman.
• Single-visit depth. Noise is a 301 s single grism exposure, not the full HLSS coadd, so the S/N distribution is pessimistic relative to the final survey.
• One healpix. All 15,434 galaxies come from healpix 10307, so large-scale environmental diversity is limited.
• Photometry is catalogue truth, noiseless as delivered. Add noise before using it; the dataset does not.

Citation

Please cite the specsr-roman repository, together with the sources of the underlying simulation:

• Troxel et al. (2025), OpenUniverse2024, arXiv:2501.05632
• Wang et al. (2022), The High Latitude Spectroscopic Survey on the Nancy Grace Roman Space Telescope, ApJ 928, 1

License

MIT, matching the code. The underlying OpenUniverse2024 products carry their own terms — see the link above.