OmniSky v1
Collection
Crossmatched, normalized multimodal astronomy: 20.4M DESI DR1 objects x 24 catalogs, 822 GB. Start with mmu-norm-index. • 20 items • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
catalog: string
repo: string
linked_payload_ids: int64
shards_done: int64
rows_delivered_from_markers: int64
split_counts: struct<val: int64>
child 0, val: int64
split_covers_all_linked: bool
ratio_delivered_over_linked: double
note: string
transform: string
n_shards_at_fit: int64
fit_on: string
n_train_rows_seen: int64
source_repo: string
stats: struct<>
to
{'catalog': Value('string'), 'source_repo': Value('string'), 'fit_on': Value('string'), 'n_train_rows_seen': Value('int64'), 'n_shards_at_fit': Value('int64'), 'note': Value('string'), 'transform': Value('string'), 'stats': {}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
catalog: string
repo: string
linked_payload_ids: int64
shards_done: int64
rows_delivered_from_markers: int64
split_counts: struct<val: int64>
child 0, val: int64
split_covers_all_linked: bool
ratio_delivered_over_linked: double
note: string
transform: string
n_shards_at_fit: int64
fit_on: string
n_train_rows_seen: int64
source_repo: string
stats: struct<>
to
{'catalog': Value('string'), 'source_repo': Value('string'), 'fit_on': Value('string'), 'n_train_rows_seen': Value('int64'), 'n_shards_at_fit': Value('int64'), 'note': Value('string'), 'transform': Value('string'), 'stats': {}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
OmniSky v1 — Index
This repository is the starting point for the OmniSky, a crossmatched astronomy dataset built around 20.4 million DESI DR1 objects. It brings together 24 source catalogs and 13 data repositories, totaling 822 GB and about 4.2 million matched objects.
Start here to find the pinned release definition, split assignments, and model-scaling statistics.
What is in this repository
| Path | Contents |
|---|---|
manifest/v1/release_manifest.json |
The release definition: repository revisions, file sizes and checksums, plus notes about how each data type was prepared |
splits/v1/spine_splits.parquet |
train/val/test assignment for every matched object |
splits/v1/payload_splits/<catalog>.parquet |
the same splits keyed by native dataset IDs, one file per source catalog |
l2_stats/v1/<catalog>.json |
robust scale parameters (median/IQR) fitted on the training split only |
qa/v1/*.json |
delivered and matched counts, plus split coverage for each data type |
How the data is organized
kshitijd/mmu-xmatch-20m contains the master object list and the catalog matches. For each source catalog, its best and candidates tables record the catalog ID, the matched MMU object ID, the separation, and any ambiguity flags. The L1 repositories (mmu-norm-*) contain normalized source data for those objects. Each dataset's notes in provenance.json describe its units, masks, repairs, and processing. The L2 repositories (mmu-l2-*) contain processed L1 data prepared for modeling, with scaling details stored alongside the data.
Splits
splits/v1 assigns each object to one split, so its spectra, images, light curves, and catalog entries stay together. The assignments use all catalog matches and a HEALPix order-4 region hash, targeting an 80/10/10 split (realized: 82.9/8.9/8.2). Splitting rows at random can reuse information about the same object across training and evaluation, especially when a value such as PROVABGS was derived from another dataset in the release. Use the supplied manifests.
Not included in v1: a temporal/prefix split for transient forecasting; peak-centered views. Build those as explicit task manifests.
Loading example
import json, pandas as pd
from huggingface_hub import hf_hub_download
m = json.load(open(hf_hub_download("kshitijd/mmu-norm-index", "manifest/v1/release_manifest.json", repo_type="dataset")))
desi = m["l1"]["desi"]
shard = hf_hub_download(desi["repo"], "shards/desi_00000.parquet", repo_type="dataset", revision=desi["revision"])
df = pd.read_parquet(shard)
Always pass the manifest's revision= value to make results reproducible.
Data notes
Each L1 repository includes data notes in provenance.json, which are also copied into the manifest. They cover approximate DESI resolution information, diagonal Gaia XP coefficient errors, missing Legacy North pixel inverse variance, Chandra net count-rate density, and dimensionless BTSbot images. Read the relevant card before using a dataset for science.
Sources and preparation
Payloads are based on the Multimodal Universe v1.0 HATS conversion (UniverseTBD), at revisions pinned per repository. This release also includes independently audited repairs: VIPERS inverse variance and redshift flags were rebuilt from ESO/CDS sources; TESS was reacquired from the TESS-SPOC HLSP at STScI with per-sector rows and QUALITY restored; and JWST weight maps were rebuilt from DAWN v7 mosaics. See the individual repository cards for upstream credits and licenses.
- Downloads last month
- 95