saturn-m

A pretraining corpus for physical time-series world models: 600.0M distinct patch tokens of multivariate sensor-shaped data, generated end to end from two synthetic processes and a seeded RNG.

No real data is read at any point in the build. No benchmark file is ever opened, so window-level leakage against any evaluation suite is structurally impossible rather than audited away. Stated precisely, because the stronger claim would be wrong: the corpus has no benchmark data contact, but it does have benchmark statistics contact — its design constants were calibrated against published measurements of real suites. That is a real, if weak, information channel from the evaluation suite into the corpus, and anyone using this data for an OOD claim should know about it.

At a glance

distinct patch tokens	600.0M (patch size 16)
windows	4,207,591
panels	2,266,149
observations	9.60B
channels per panel	C ∈ {1, 2, 4, 7, 14, 25, 43, 64}, mean 7.67
window length	≤ 512 samples (T ∈ {128, 256, 384, 512}, plus irregular decimations)
sampling rates	1/day → 20 kHz, nine decades, flat by token share
dynamical families	11, effective 10.94 by token share
files	108 parquet shards under data/, per-shard manifests and cards under shards/
seed	3400–3405 (one per build shard)

Schema

One row is one PANEL, not one channel. target is list<list<float32>> with shape (n_channels, length), so the multivariate grouping survives any reordering of the rows. This is deliberately not the ['start','target'] layout in which grouping is recoverable only as index // F.

column	meaning
target	list<list<float32>>, shape (n_channels, length) — channel-major
channel_index	explicit 0…C−1
group_id	{corpus}:{seed} — names the generating panel
n_channels, length	panel shape
backbone	gen_broad (eleven dynamical families) or cauker (causal DAG over GP nodes)
family	one of the eleven families below
obs_model	clean / quantised / heavy_tail / missing / regime
erank_target	the requested effective-rank fraction (a stratum coordinate)
corr_erank_frac	the measured effective rank / C; null for C=1
L_gen	length actually generated before windowing (512 / 1024 / 2048)
rate_hz	nominal sampling rate — the seconds-to-years axis
irregular	whether the window was non-uniformly decimated
acf1_target	target lag-1 autocorrelation (0.3 / 0.6 / 0.9 / 0.97)
seed, window_in_panel	regeneration coordinates
K	latent bases (gen_broad); −1 for CauKer
max_parents	DAG density (CauKer); −1 for gen_broad

Generators

gen_broad supplies the eleven dynamical families. It is a coverage generator, not a physics simulator: each family exists to close one measured feature-coverage failure against a real benchmark, and the mixture is forced by build gate 9 rather than chosen by taste.

family	what it is	the gap it closes
longmem	fractionally-integrated noise	an ACF that decays as a power law, not a geometric one. An AR bank that still has memory at lag 32 is forced to acf1 ≈ 0.99 and then fails acf1's own range.
degradation	monotone drift to failure	trend_r2 near 1 — the C-MAPSS / BATADAL shape. A random walk only reaches trend_r2 0.44 at p95; the degradation benchmarks sit above it.
gp_kernel	KernelSynth	1–3 kernels composed from a bank with + and ×, GP sampled once. The Chronos-style backbone, kept as one family among eleven rather than the whole corpus.
arma	AR(p) with sampled root magnitude	sweeps acf1 continuously from ~0 (white) to ~0.99 (near unit root).
walk	integrated noise, optionally with drift	high trend_r2 at low diff_sd_ratio.
seasonal	trend + 1–3 seasonal components + noise	the classic decomposition over a wide period range.
regime	Markov regime switching	discrete level shifts, fat tails, low spectral entropy — a GP is stationary and cannot do this.
piecewise	piecewise-constant	the frac_repeat / n_distinct axis no GP can reach.
chaotic	logistic / tent map	deterministic, low spectral entropy, non-Gaussian marginal.
spiky	sparse point process on a quiet baseline	the kurtosis 469 / skew 15.7 corner measured on SMAP.
smooth	twice-integrated / low-pass noise	the ultra-smooth corner GIFT-Eval has and the 13 benchmarks do not (covid_deaths and bizitobs_l2c_5T both score 1/6 on TEMPORAL, and miss the same way).

CauKer (ed006705ca, ShifengXIE/CauKer) supplies a causal DAG over GP-valued nodes, confined to panels of ≥ 14 channels — below that it has no cross-channel structure to contribute and its cost per token is up to 84× higher. It carries 15.5% of the corpus's tokens.

The observation model is a labelled post-process applied on top of the dynamics — it is what turns a process into a sensor reading, and supplies what a Gaussian process cannot:

obs_model	what it does
clean	the process is passed through untouched.
quantised	each channel is rounded onto a random 2–48 level grid. SMAP's real sensors repeat 97% of their values over ~1.8 distinct levels and no Gaussian process does that.
heavy_tail	a signed power raises the tail to kurtosis > 100 and leaves skew ≈ 0.
missing	two real sensor faults, both bursty rather than i.i.d.: hold-last-value (a stuck transmitter) and hard dropout to zero (a dead channel). Realised in the values — there is no NaN and no validity mask, because the pack format carries neither.
regime	2–5 segments with level shifts and variance drift, since a GP is stationary by construction.

Note that gen_broad also applies its own random, unlabelled behaviours internally; this layer sits on top and is the one the obs_model column names.

Build gates

All ten pass. Each is a measurement with a pre-registered threshold, not a description.

#	gate	threshold	measured
1	dedup	0 duplicate windows	0	PASS
2	SD floor	< 2% of windows below 1% of corpus SD	0.56%	PASS
3	panel coupling	one group_id per PANEL, never index // F	2,266,149 panels over 4,207,591 windows	PASS
4	rank calibration	non-zero mass below corr_erank_frac 0.11, spanning to 1.00	1.74% below 0.11, support [0.016, 1.00]	PASS
5	clock check	event probe from window position alone stays at its permutation null	excess -0.031 (mean AUROC 0.471 vs null 0.501)	PASS
6	leakage	zero real benchmark windows	0	PASS
7	coverage	>= 85% of benchmark features, CROSS >= 86%	94.4% overall (204/216 cells), CROSS 93.1%	PASS
8	channel balance	>= 6 effective C levels by token share	6.41	PASS
9	family balance	>= 10.5 effective families by token share	10.94	PASS
10	cross-channel coupling	differenced mean |r| above its analytic null of 0.0353	0.212 (6.0x the null)	PASS

Loading

from datasets import load_dataset
import numpy as np

ds = load_dataset("Forgis/saturn-m", split="train", streaming=True)
row = next(iter(ds))
x = np.array(row["target"], dtype=np.float32)   # (n_channels, length)
assert x.shape == (row["n_channels"], row["length"])

The manifest carries every stratum coordinate for every window without touching the parquet:

from huggingface_hub import hf_hub_download
import numpy as np
m = np.load(hf_hub_download("Forgis/saturn-m", "shards/sh00_manifest.npz",
                            repo_type="dataset"), allow_pickle=True)
m["family"], m["C"], m["corr_erank_frac"]

Caveats

• Coverage is measured marginally — a benchmark median inside the corpus p5–p95, one feature at a time. It is an upper bound, never "94.4% of the benchmark distribution is reproduced".
• Missingness is in the values, not in a mask. obs_model="missing" produces held and zeroed samples; there is no NaN and no validity channel to read it off.
• Irregular sampling has no timestamp axis. It is realised as non-uniform decimation, so a model sees a varying effective Δt but never a clock.
• Window length is capped at 512 by the evaluation context. The scale axis is carried by rate_hz, not by sample count.
• Channel counts take eight discrete values, not all of 1…64. The tokenizer is channel-agnostic (one shared Linear(P, d) across channels, no channel embedding), so a test system with an unseen C is interpolation rather than extrapolation — measured flat across the ladder, with no penalty at untrained values.
• Built as six independent shards, deduplicated against one global hash set carried shard to shard, so the 0-duplicate gate holds across the whole corpus and not merely within a shard.

Reproducibility

Seed 3400–3405 (one per build shard), CauKer commit ed006705cabbc5f0957da6de945e911cda30a387. Every panel is regenerable from its recorded stratum coordinate and seed.

Related

• Forgis/saturn-s — the smaller corpus, 1/6.1 the tokens from the same build grid, passing the same ten gates. The channel ladder and the family mixture match to within 0.1 effective levels, so the data axis is close to held fixed across the pair — with one known exception: irregular sampling reaches 13.7% of saturn-s's tokens against 19.1% of saturn-m's, both short of the 20% design target. Treat irregular as the one axis that is not matched between the two.