EiffL/aion2-full-lancedb

AION-2 DESI and Legacy Survey LanceDB

Two complete survey tables containing existing AION-1 tokens, prepared for AION-2 training and live crossmatching. The native LanceDB database occupies approximately 10.32 GB (9.61 GiB), including coordinate indexes. Raw spectra and image pixels are not included.

Table / configuration	Observations	Modality column	Tokens per row	Vocabulary	Lance table version
desi	1,126,441	tok_spectrum_desi	273	1,024	3
legacysurvey	14,174,203	tok_image	576	4,375	3

Token columns retain their native {token: [...]} structs and integer IDs. Original survey metadata, including ra, dec, object_id, and _healpix_29, is retained. source_partition and source_row identify the source Parquet file and row. Every input row is included. Each table has native BTree indexes on ra and dec for live spatial matching; no precomputed crossmatch table or pair index is included.

Download for training

pip install -U huggingface_hub lancedb pylance
hf download EiffL/aion2-full-lancedb --repo-type dataset --local-dir data/aion2-hf

Then open the downloaded database:

import lancedb

db = lancedb.connect("data/aion2-hf/full.lancedb")
desi = db.open_table("desi")
legacy = db.open_table("legacysurvey")
desi.checkout(3)
legacy.checkout(3)
print(desi.count_rows(), legacy.count_rows())

In an AION-2 training recipe, set config.dataloader.dataset = "data/aion2-hf/full.lancedb" and pin both SurveyConfig entries to version 3. The model registry selects tok_image and/or tok_spectrum_desi; mixture weights and live matching rules remain part of the training configuration.

For reproducible downloads, add --revision <Hub-commit-SHA> to hf download. The Hub commit identifies this repository snapshot; Lance table version 3 identifies the indexed snapshot within each table. Retain both when recording a training run. Training reads should use the local copy.

Query directly from Hugging Face

Hugging Face supports native Lance tables, including bundled indexes. For small exploratory queries, open the database without downloading it all:

import lancedb

db = lancedb.connect("hf://datasets/EiffL/aion2-full-lancedb/full.lancedb")
table = db.open_table("desi")
table.checkout(3)
rows = table.search().select(["object_id", "ra", "dec"]).limit(5).to_list()
print(rows)

The Hugging Face datasets API also supports streaming each survey separately:

from datasets import load_dataset

spectra = load_dataset(
    "EiffL/aion2-full-lancedb", "desi", split="train", streaming=True
)
first = next(iter(spectra))
print(len(first["tok_spectrum_desi"]["token"]))  # 273
# Use configuration "legacysurvey" for image tokens.

These examples use the current Hub revision. For pinned training snapshots, use the CLI download above and native LanceDB. Repeated remote random-access queries can be throttled by Hub rate limits; the local database supports the same projected reads and live crossmatches without that network overhead. See Hugging Face's Lance guide and LanceDB's upload guide.

Provenance and validation

Source dataset	Pinned revision
DESI AION tokens	bba0be0d28724d9fee38dae2cff034993ca2d6ac
Legacy Survey AION tokens	b8707333e3b25addae7faedf28eb076b0ae1eda5

The upstream token catalogs were produced with aion-hats, using the polymathic-ai/aion-base codec, from UniverseTBD/mmu_desi_edr_sv3 and UniverseTBD/mmu_ssl_legacysurvey_north. This repository repackages their existing tokens into Lance; it does not retokenize or generate observations. The source cards do not declare a license, so this conversion does not assert a new data license; consult the original survey and token-dataset terms and attribution requirements.

All 5,794 source Parquet files were checked against their published SHA-256 hashes. Ingestion validated row counts, token lengths, vocabulary bounds, and coordinates. Both tables have zero null token rows and zero invalid coordinates. Sampled token/metadata values were checked against the original Parquet files, and sampled live matches were independently checked by angular separation.

full.lancedb/manifest.json records the source revisions, partition hashes, row counts, and table versions. SHA256SUMS records the published database file checksums. The database was built with LanceDB 0.39.0, Lance 12.0.0, and PyArrow 25.0.1. Keep the table directories intact, including _versions, _transactions, _indices, and data.