MaxRonce/aion2-six-survey-continuous-lancedb

AION-2 Sep-run LanceDB

Continuous astronomical tokens from the September 2026 AION-2 experiments, exported as a LanceDB database compatible with the September AION-2 clean reader. The export contains 173,398,335 catalogue rows, 8 physical tables, and 18 modality columns across six survey families: COSMOS, Euclid, HSC, Legacy Survey, DESI, and SDSS.

The dataset combines three different kinds of representation:

• Images: a learned GalactiTok ImageLoLa convolutional autoencoder, with an 8 x 8 spatial bottleneck and 6 latent channels.
• Spectra: Jeff Shen's SpectrumJeff / Universal Spectrum Tokenizer, a wavelength-aware Transformer autoencoder with 6-dimensional continuous patch tokens and a leading normalization token.
• Photometry and scalar redshifts: GalactiTok deterministic continuous codecs, using fitted catalogue normalization statistics. These are not learned image or spectrum autoencoders.

All stored token coordinates are float32. They are not integer codebook indices, textual tokens, or embeddings produced by a trained AION Transformer. The common width of six does not imply that the different tokenizers share coordinate meanings, scales, or a jointly aligned latent space.

The repository identifier remains MaxRonce/aion2-six-survey-continuous-lancedb; AION-2 Sep-run LanceDB is the display name. The database directory is six-survey-continuous-full.lancedb/.

Dataset inventory

HSC Deep/Wide and Legacy Survey North/South are separate tables. The historical name "six-survey" refers to survey families, not the number of Lance tables. Rows are survey catalogue entries, not 173 million distinct cross-survey objects: the same astrophysical source can occur in several tables.

Lance table	Rows	Image tokens	Spectrum tokens	Photometry tokens	Redshift tokens
cosmos	484,487	64 x 6	-	5 x 6	-
euclid_q1	29,953,430	64 x 6	-	4 x 6	-
hsc_deep	474,954	64 x 6	-	5 x 6	-
hsc_wide	3,219,412	64 x 6	-	5 x 6	-
legacysurvey_south	123,159,232	64 x 6	-	8 x 6	-
legacysurvey_north	14,174,203	64 x 6	-	3 x 6	-
desi	1,126,441	-	245 x 6	3 x 6	1 x 6
sdss	806,176	-	147 x 6	5 x 6	1 x 6
Total	173,398,335	6 modalities	2 modalities	8 modalities	2 modalities

The source export occupied approximately 294 GiB on Jean-Zay. Downloaded size, local allocated space, and Hub-reported storage can differ.

What one token means

Throughout this card, a shape N x 6 means N tokens, each containing six real numbers. Lance stores these tensors as a one-dimensional float list per row. For a modality with N tokens:

tokens = np.asarray(row[column], dtype=np.float32).reshape(N, 6)

Flattening changes the storage layout, not the information or precision of the retained token values. It is unrelated to the coordinate indexes.

For images, the tensor is 64 x 6 for the whole multiband image, not 64 x 6 x number_of_bands. The six latent coordinates are learned features, not six physical filters. Photometry is different: it explicitly stores one six-coordinate token per configured band.

Complete modality schema

Every table also contains ra: float64, dec: float64, and object_id: string. RA and Dec are sky coordinates in degrees. Object IDs should be interpreted within their survey, not as a universal cross-survey key.

Table	Token column	Logical shape	Float32 values per non-null row	Token ordering
cosmos	image_cosmos	64 x 6	384	Row-major 8 x 8 latent grid
cosmos	photometry_cosmos	5 x 6	30	F115W, F150W, F277W, F444W, F770W
euclid_q1	image_euclid_q1	64 x 6	384	Row-major 8 x 8 latent grid
euclid_q1	photometry_euclid_q1	4 x 6	24	VIS, Y, J, H
hsc_deep	image_hsc_deep	64 x 6	384	Row-major 8 x 8 latent grid
hsc_deep	photometry_hsc_deep	5 x 6	30	g, r, i, z, y
hsc_wide	image_hsc_wide	64 x 6	384	Row-major 8 x 8 latent grid
hsc_wide	photometry_hsc_wide	5 x 6	30	g, r, i, z, y
legacysurvey_south	image_legacysurvey_south	64 x 6	384	Row-major 8 x 8 latent grid
legacysurvey_south	photometry_legacysurvey_south	8 x 6	48	g, r, i, z, W1, W2, W3, W4
legacysurvey_north	image_legacysurvey_north	64 x 6	384	Row-major 8 x 8 latent grid
legacysurvey_north	photometry_legacysurvey_north	3 x 6	18	g, r, z
desi	spectrum_desi	245 x 6	1,470	1 normalization token + 244 spectral patches
desi	photometry_desi	3 x 6	18	g, r, z
desi	spectroscopic_redshift_desi	1 x 6	6	One normalized catalogue Z value
sdss	spectrum_sdss	147 x 6	882	1 normalization token + 146 spectral patches
sdss	photometry_sdss	5 x 6	30	u, g, r, i, z
sdss	spectroscopic_redshift_sdss	1 x 6	6	One normalized catalogue Z value

Image, photometry, and redshift columns are Arrow large_list<float32>. The spectral columns are fixed_size_list<float32>[1470] for DESI and fixed_size_list<float32>[882] for SDSS. Photometry columns can be null; see the missing-data policy below. Shapes and tokenizer recipes must not be inferred from the Arrow list type alone.

Image tokenizer: GalactiTok ImageLoLa

Model identity

All six image modalities use the same multisurvey ImageLoLa checkpoint, identified in the production provenance as image-lola-six-300m-8x8x6-compare-1990706/final-pretrained.

The provider class is named ImageVQVAETokenizer, but this configuration has no quantizer (quantizer_config: null). It uses ImageLoLaEncoder and ImageLoLaDecoder, adapted from the LoLa deep-compression autoencoder associated with Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation. The class name does not mean that these stored tokens are VQ codes.

Encoding architecture

The materialization adapter uses 160 x 160 image crops. Available filters are placed in the checkpoint's ordered global band axis and accompanied by a channel-availability mask. A masked per-pixel linear projection combines the available filters into a shared feature image before the convolutional encoder. The input projection masks absent channels and scales its output by sqrt(number_of_global_bands / number_of_present_bands).

The corresponding six-survey 8 x 8 source recipe specifies:

Component	Recipe configuration
Input spatial size	160 x 160
Multisurvey projection width	80 features per pixel
Encoder stage widths	112, 224, 448, 880
Residual blocks per stage	3, 3, 3, 3
Spatial transition factors	2, 2, 5
Spatial sizes	160 x 160 -> 80 x 80 -> 40 x 40 -> 8 x 8
Residual block	Channel normalization, 3 x 3 convolution, SiLU, dropout, 3 x 3 convolution, residual addition
Downsampling	PixelUnshuffle / space-to-depth followed by convolution
Latent projection	3 x 3 convolution to 6 channels
Latent saturation	z / sqrt(1 + (z / 5)^2)
Decoder	Mirrored stages with convolution and PixelShuffle upsampling
Quantization	None

The source recipe is described here separately from serialized-checkpoint verification: the exported checkpoint's config.json is authoritative for exact architecture and calibration settings, and is not bundled in this dataset repository. The checkpoint is from the approximately 300M-parameter image-tokenizer experiment; this is not the parameter count of the AION model that later consumes the tokens.

The convolutional encoder jointly compresses the multiband image to (6, 8, 8). The AION adapter changes this to (64, 6) using latent.flatten(2).transpose(1, 2), preserving the spatial order. A row then stores the flattened 384 float32 values. Latent cells have convolutional receptive fields; they are not independently encoded raw 20 x 20 image patches.

Image bands versus photometric bands

Image table	Image filters supported by the six-survey recipe
COSMOS	F115W, F150W, F277W, F444W, F770W
Euclid Q1	VIS, NIR Y, NIR J, NIR H
HSC Deep and HSC Wide	HSC g, r, i, z, y
Legacy Survey South	g, r, i, z
Legacy Survey North	BASS g, BASS r, MzLS z

These filters occupy survey-qualified channels in the multisurvey model. Identically named optical filters are not assumed to have identical response functions. The availability mask identifies which channels a sample supplies. Legacy South's W1-W4 photometry does not imply W1-W4 image channels in this image tokenizer.

Image normalization and decoding

ImageLoLa uses MMOMAImageNormalization: survey/band zero-point and pixel-scale calibration, checkpoint-configured clipping where enabled, and an asinh dynamic range compression before the encoder. The source template uses reference AB zero point 22.5, reference pixel scale 0.262 arcsec/pixel, compression factor 0.01, and multiplier 10. Actual per-band calibration and clipping values belong to the checkpoint, not to the six latent coordinates.

With saturation bound 5, the encoder's final transform is bounded by approximately [-5, 5], not [-1, 1]. It is not a min-max normalization of the catalogue.

To reconstruct images, the float lists alone are insufficient: use the matching decoder weights, normalization configuration, band order, availability context, and survey pixel scale. The compact database does not carry original images, image inverse-variance maps, or all image calibration metadata.

Spectrum tokenizer: Jeff Shen / SpectrumJeff

DESI and SDSS use Jeff Shen's Universal Spectrum Tokenizer, exposed by GalactiTok as SpectrumJeffTokenizer:

• Public checkpoint: j-shen/universal_spectrum_tokenizer.
• Pinned checkpoint revision: f903bc6aed0e38367a61eff317dc97bb28d06066.
• Serialized configuration at that revision.

Architecture and sequence layout

Component	Checkpoint configuration
Wavelength samples per patch	32
Transformer hidden width	384
Attention heads	8
Encoder Transformer blocks	6
Decoder Transformer blocks	2
MLP ratio	4
Dropout setting	0.1; inactive in evaluation mode
Bottleneck width	6 continuous coordinates per patch
Reconstruction channels	2: flux and a bad-pixel-mask prediction
Wavelength embedding	Sinusoidal, periods 0.001 to 20,000
Input inverse-variance cap	10,000
Encoder masking fraction in checkpoint config	0.0

The encoder takes flux, wavelength, and inverse variance. Its implementation builds two input channels from transformed flux and inverse standard deviation, groups them into 32-pixel patches, applies a linear patch embedding, and adds sinusoidal wavelength embeddings. Transformer blocks exchange information between patches before a linear projection reduces each patch to six values. Patch tokens are therefore contextual representations, not independent local summaries. The blocks use attention with normalized queries/keys and a gated feed-forward network.

One extra token is prepended to preserve the per-spectrum flux normalization:

• DESI: 1 + 244 = 245 tokens, representing a patch-aligned length of 244 x 32 = 7,808 wavelength samples.
• SDSS: 1 + 146 = 147 tokens, representing a patch-aligned length of 146 x 32 = 4,672 wavelength samples.

These lengths describe the tokenizer input after any reader padding. They do not imply that every original spectrum has that many valid measured pixels. The leading token is not a class token or a redshift token: it stores the normalization scalar repeated across its six coordinates. The subsequent coordinates are learned spectral features. Do not drop or permute the prefix when decoding.

Normalization, masks, and reconstruction

The provider derives bad pixels from ivar < 1e-6, computes a positive flux normalization from usable pixels, stores its logarithmic representation in the prefix token, and applies asinh transformations to the patch inputs. Patches with at least half their pixels marked bad receive a learned mask token. The AION adapter also maps explicitly invalid input pixels to zero inverse variance before calling the provider.

The decoder projects the six-dimensional patch tokens back to width 384, adds the target wavelength embeddings, applies its two Transformer blocks, and reconstructs flux and mask channels. The provider's training loss combines an inverse-variance-weighted flux reconstruction term with a bad-pixel-mask term.

There is no discrete codebook or guaranteed [-1, 1] range for these spectral tokens. The prefix and the learned patch coordinates have different roles and can have different distributions. The stored spectra are not lossless raw spectra; wavelength grids and original uncertainty arrays are not preserved as separate columns in this compact export. Decoding requires the matching checkpoint and the correct wavelength grid/padding convention. The decoder does not reconstruct the original inverse variances.

Photometry tokenizer: deterministic six-coordinate features

The eight photometry modalities use GalactiTok's PhotometryTokenizer with token_dim=6. This codec has no learned neural encoder or decoder. Its survey-specific artifacts contain ordered bands and fitted location/scale statistics. Each band is encoded independently, with one token per band.

For a band, let f be flux, w its inverse variance, m the validity mask, and location, scale the fitted normalization constants. Define:

u = (f - location) / scale
q = w * scale^2

token[0] = u
token[1] = asinh(q)
token[2] = asinh(u)
token[3] = asinh(u * sqrt(q))
token[4] = log1p(q)
token[5] = float(m)

Thus six coordinates do not mean six independent measured quantities. Coordinates 0/2 describe the same normalized flux with different dynamic ranges; 1/4 describe normalized uncertainty; 3 combines flux and uncertainty; 5 records validity. The codec is redundant by design.

Decoding uses f = token[0] * scale + location and w = sinh(token[1]) / scale^2. The other coordinates are not necessary to invert the original valid measurement. For tokens generated by a model, consistency between these redundant coordinates is not guaranteed automatically.

Fitting and input measurement conventions

The fitting script streams each source catalogue and uses a bounded random reservoir of finite valid measurements per quantity. It fits the location with the median and the scale with 1.4826 * median(abs(f - median(f))), falling back to standard deviation and then 1 if the scale is degenerate. Statistics are stored in float32. This is catalogue-statistics fitting, not gradient-based autoencoder training.

Table	Catalogue measurements used	Uncertainty input
cosmos	MAG_MODEL_F115W, MAG_MODEL_F150W, MAG_MODEL_F277W, MAG_MODEL_F444W, MAG_MODEL_F770W	Not supplied; explicit zero inverse variance
euclid_q1	mer_catalogue_flux_{vis,y,j,h}_2fwhm_aper	Corresponding mer_catalogue_fluxerr_*_2fwhm_aper standard deviations
hsc_deep, hsc_wide	{g,r,i,z,y}_cmodel_mag	Corresponding *_cmodel_magerr magnitude errors
legacysurvey_south	FLUX_{G,R,I,Z,W1,W2,W3,W4}	Not supplied; explicit zero inverse variance
legacysurvey_north	flux_g, flux_r, flux_z	Not supplied; explicit zero inverse variance
desi	FLUX_G, FLUX_R, FLUX_Z	FLUX_IVAR_G, FLUX_IVAR_R, FLUX_IVAR_Z
sdss	SPECTROFLUX_{U,G,R,I,Z}	SPECTROFLUX_IVAR_{U,G,R,I,Z}

The AB-magnitude reader converts magnitudes to flux with zero point 22.5: f = 10**(0.4 * (22.5 - magnitude)). Supplied magnitude uncertainties are propagated into flux uncertainty. Native flux columns follow their source catalogue convention; do not assume that all surveys provide the same units, apertures, or photometric estimator merely because the output width is six.

The artifact family is named six-survey-scalar-tokenizers-6d-zero-ivar-v1. "zero-ivar" does not mean that all eight surveys lost their uncertainties. It allows sources without error columns to retain finite measurements with explicitly unknown precision. For these valid zero-ivar measurements, the token is [u, 0, asinh(u), 0, 0, 1]. Zero inverse variance is not zero measurement error and does not, by itself, make the flux invalid.

Neither the linear normalized flux nor its asinh transform has a fixed [-1, 1] or [0, 1] range. Bright or unusual objects can produce very large finite coordinates. Such values must not be confused with discrete token IDs or silently clipped without defining a new data contract.

Spectroscopic-redshift tokenizer: normalized scalar with padding

DESI and SDSS each have one redshift modality, derived from their catalogue Z column. The provider is GalactiTok's ContinuousScalarTokenizer, called through the AION ScalarTokenizer adapter, with an AffineScalarNormalization and token_dim=6.

u = (Z - location) / scale
token = [u, 0, 0, 0, 0, 0]
Z = token[0] * scale + location

Location and scale use the same robust fitting procedure as above, but are fitted separately for DESI and SDSS. Only the first coordinate carries the scalar; the other five are zero padding for the shared six-coordinate storage contract. This is not a learned six-dimensional redshift embedding, a redshift probability distribution, or six redshift measurements. ZERR and ZWARN are not part of this token or separate columns in the migrated export.

The physical redshift can be recovered up to floating-point roundoff only with the correct survey's fitted location and scale. These constants are not recoverable from the six-coordinate shape alone and are not included in the current dataset files.

Migration and missing-data policy

This release is a schema translation, not a fresh tokenization. It was built from version 1 of the September multimodal token database. Coordinate and ID columns were renamed to the clean convention; token columns were renamed to modality names; retained float values were copied; and new scalar BTree indexes were built on ra and dec. Each output table is at Lance version 3.

The source database carried band-level photometry validity masks. The compact clean representation uses modality-level nullability instead:

• All bands valid: retain the entire photometry token list.
• Some bands invalid: set the whole photometry column to null for that row.
• No bands valid: set the whole photometry column to null.
• Keep the row and its other modalities in all three cases.

Consequently, a source row with partial photometry is not represented by a shorter token sequence or a zero-filled valid measurement. Some otherwise-valid bands are intentionally unavailable in this clean export. The original band-level mask column is not retained.

Table	All bands valid; photometry retained	Partially valid; photometry null	No valid bands; photometry null
cosmos	484,487	0	0
euclid_q1	29,277,096	593,674	82,660
hsc_deep	472,943	2,011	0
hsc_wide	3,217,158	2,254	0
legacysurvey_south	123,159,232	0	0
legacysurvey_north	14,174,203	0	0
desi	1,042,509	0	83,932
sdss	806,176	0	0

The full migration receipt reports zero incorrect token lengths and zero non-finite retained token values. These are structural checks, not evidence of reconstruction fidelity, astrophysical correctness, or independent train/test splits. Invalid photometry was nulled before the finite-value audit.

LanceDB storage and live crossmatching

six-survey-continuous-full.lancedb/
  cosmos.lance/
  euclid_q1.lance/
  hsc_deep.lance/
  hsc_wide.lance/
  legacysurvey_south.lance/
  legacysurvey_north.lance/
  desi.lance/
  sdss.lance/
  MIGRATION_COMPLETE.json
  MIGRATION_COMPLETE.sha256

Each table directory includes data fragments, version manifests, transaction metadata, and its RA/Dec BTree index files. Preserve the complete directory layout. Individual .lance data files alone are not an openable Lance table.

The indexes accelerate scalar coordinate filtering for live crossmatching. They are not latent-vector indexes, and they do not encode a precomputed table of matched objects. The consuming application must still define its anchor, target, angular search radius, angular-distance check, and match policy. A one-arcsecond radius was used in the September clean compatibility tests; it is a reader configuration, not an intrinsic property of the database.

No crossmatch Parquet indexes, fixed training mixture, sampling probabilities, or train/validation/test partitions are bundled. The database can supply both single-survey and crossmatched examples, depending on the training recipe.

Download and read

Download the entire database folder from one immutable Hub commit. Select a complete upload revision, then pin its SHA for reproducibility:

export HF_REVISION="YOUR_VERIFIED_COMPLETE_COMMIT_SHA"
hf download MaxRonce/aion2-six-survey-continuous-lancedb \
  --repo-type dataset --revision "$HF_REVISION" \
  --include 'six-survey-continuous-full.lancedb/**' \
  --local-dir ./aion2-sep-run

This is a native LanceDB export, not a conventional Hugging Face Parquet dataset for datasets.load_dataset. Read it with LanceDB:

from pathlib import Path

import lancedb
import numpy as np

path = Path("aion2-sep-run/six-survey-continuous-full.lancedb")
if not (path / "desi.lance").is_dir():
    raise FileNotFoundError("Download the complete LanceDB directory first")
db = lancedb.connect(str(path))
table = db.open_table("desi")
table.checkout(3)
row = table.search().select([
    "object_id", "ra", "dec", "spectrum_desi", "photometry_desi"
]).limit(1).to_list()[0]

spectrum = np.asarray(row["spectrum_desi"], dtype=np.float32).reshape(245, 6)
normalization_token = spectrum[0]
spectral_patch_tokens = spectrum[1:]
photometry = (
    None if row["photometry_desi"] is None
    else np.asarray(row["photometry_desi"], dtype=np.float32).reshape(3, 6)
)
print(table.count_rows(), spectrum.shape)

For an image row, reshape the float list to (64, 6). Recover the original channel-first latent layout with tokens.T.reshape(6, 8, 8) before passing it to the appropriate decoder.

For the September clean recipe, modality names must match the column names above, with modality_type="continuous", native_dim=6, and the corresponding num_tokens. Compatibility was tested against clean commit aa2e06f. This does not promise compatibility with every later AION API or configuration schema. AION model-level target standardization, projections, and flow heads are downstream choices and are not baked into these stored tokens.

Provenance and verification scope

• Source token database: six-survey-multimodal-tokens-8x8x6-v1.lancedb, version 1.
• Clean export: six-survey-continuous-full-20260924-101901.lancedb.
• Full migration elapsed time in the receipt: approximately 1,218 seconds.
• Receipt: MIGRATION_COMPLETE.json.
• Receipt SHA-256: 15ff6bbbfa5a2e4d9f01378f2eca5346a61658c7f6e43cd31d726fc3faa09d7e.

The September 29 Hub inventory at commit 85d09e9b3d323778e21e6b6b43f8d88a6c94bb9a includes all eight table directories, including 118 Legacy South data fragments plus its 11 metadata/index files. The receipt hash was verified. This inventory check is not a full remote scan or a checksum verification of every data fragment. The migration receipt describes the source export; its presence alone does not prove an upload or download completed. The original .sha256 sidecar includes an absolute Jean-Zay path, which must be adjusted before using sha256sum -c elsewhere.

Production evaluation provenance records the following image checkpoint identifiers; these are supplied for identification, not as downloadable weights:

ImageLoLa final-pretrained/config.json SHA-256:
7aea015e44d12d81195d9ee7885aab0820a427bcf509aea269f4776f1eb2d361
ImageLoLa final-pretrained/model.safetensors SHA-256:
786e8f94a062392ace46b2e79dbbf8609a5874622b68b023bab7c3933640decc
SpectrumJeff config.json SHA-256:
35222b0a724ce6350d14c4b0751ebf67236c8aca1a40607c5d553f2a1d6eb405

The descriptions were checked against the production table/token contracts, the scalar diagnostics, the published SpectrumJeff configuration, and these implementation paths in the corresponding development checkouts:

• GalactiTok models/image_vqvae/image_lola.py, subsampler.py, normalization.py, and scripts/prepare_six_survey_training.py for images.
• GalactiTok models/spectrum_jeff/modeling_spectrum_jeff.py for spectra.
• GalactiTok models/photometry/modeling_photometry.py and models/scalar/modeling_scalar.py for the deterministic codecs.
• AION scripts/plan_six_survey_scalar_tokenization.py and scripts/fit_six_survey_scalar_tokenizers.py for measurement contracts and normalization fitting.
• AION scripts/translate_continuous_six_surveys.py for this clean export.

The source references explain the architecture but do not replace the exact tokenizer artifacts needed for reproducible decoding. The image checkpoint and the survey-specific photometry/redshift config.json and fitting receipts must be obtained separately; they are not included in the current upload.

Intended use and limitations

• Intended for continuous-token multimodal learning, live-crossmatch loading experiments, and controlled comparisons using the documented representation.
• This is not an independently calibrated scientific catalogue or a release of all raw images, spectra, masks, wavelengths, and photometric metadata.
• Learned image/spectrum compression can lose information. A shared width of six is a software interface, not a guarantee of equal reconstruction quality or survey invariance.
• Some token channels are redundant or constant. Pooling all six coordinates into one range or variance statistic can hide their different meanings.
• Null photometry must remain unavailable; do not substitute zeros and treat them as genuine measurements.
• Create and audit your own leakage-safe splits. Catalogue rows, crossmatched counterparts, and overlapping survey observations are not independent by default. Normalization fitting here is not a certification of held-out-data isolation.
• Include tokenizer identity, preprocessing, valid-modality policy, dataset revision, crossmatch policy, and sampling mixture when comparing experiments.

Licensing and attribution

The metadata uses license: other because a single verified redistribution license for the combined derived corpus has not been specified here. This card does not grant additional rights over the original surveys or tokenizer weights. Consult the relevant source-data and model terms before redistribution or reuse.

Attribute the source surveys, AION-2, GalactiTok, Jeff Shen's Universal Spectrum Tokenizer, and the LoLa architecture as applicable. Tokenizer implementation licenses and dataset rights are separate; a code license does not automatically license the underlying astronomical data.