Datasets:

graziul
/

tokamark-artifacts

Languages:

English

ArXiv:

Tags:

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

TokaMark Working Artifacts

Documentation set supporting a contribution to the TokaMark benchmark: diagnostics, errata, specifications, and governance.

Version: 0.5

STATE UPDATE 2026-09-26 (v0.5) — scope and provenance. Access is not a contribution: the corpus is anonymously readable on S3 (11,574 shots) and the upstream repo already ships a windower, split CSVs and an evaluator. This repository contributes a provenance lock, a measured frozen split, a protocol-matched metric module, and a harness — with the harness gating all model work. The metric protocol is documented in EVALUATION-PROTOCOL.md. The findings table below is current through X60; X61-X72 are in GOVERNANCE.md Section 5. Status: working draft. Research notes, not peer-reviewed claims. Every measured quantity is tagged and traceable. Licence: CC BY-SA 4.0, matching the upstream provenance (X9). Canonical ledger: GOVERNANCE.md §5. Identifier cross-index: GOVERNANCE.md §4.5.

Read this first

Four headline claims this set previously carried are withdrawn, downgraded or reversed. They are listed here because a reader who finds a stale copy needs to know which numbers are gone.

"Error is core-concentrated by 12.47x" — WITHDRAWN. No candidate mask reproduces it from the stored residual map, and the mask was never recorded. The binned profile is ~2.06x raw and ~1.20x after normalising by local σ, which is U-shaped and worst at the extreme edge. (X30, X31) [X30]

"TokaMind halves the CNN." False. The CNN's Task 1-3 value is 0.1481, not 0.255 — a number appearing in neither paper. The gain is 17.6%. (X18)

"There is no streaming path." False. The toolkit documents S3 + fsspec and ships an IterableDataset. What is missing is Hub presentation. (X22)

"The truncation floor is ~0.0427 at k=7." The correction was the error: at k=7 the value is 0.0913, and 0.0427 belongs to k≈9–11. (X40) [X40]

Every correction is in GOVERNANCE.md §5 with its evidence. The ledger is append-only and now stands at 52 X-IDs in two days; every recent addition is a correction to our own work — including two about this document set's own compliance: a governance revision that asserted a version-header state it had not checked (X51), and two identifier prefixes that had been colliding in our own prose (X52).

Why this exists

TokaMark is a 14-task benchmark over 11,573 MAST shots (58 archives, 565 GB).

Access. The Hub viewer cannot render the archive (0 configs, 0 rows, 1 failed parquet job). (A documented S3 path exists — X22 — but it is not visible from the card.)
Evaluation — the part that survived. The flux-map task (1-3) has no physics-consistency metric, and its aggregate NRMSE is tail-dominated and variance-dominated:
- the top 1% of windows carry 47.9% of test MSE (top 5%: 60.3%); ~18 of 1,755 windows decide a comparison (X36);
- corr(per-pixel RMSE, per-pixel target σ) = 0.970; normalising by local σ collapses the spread from 2.06x to 1.20x (X31). So a single NRMSE largely reports the field's own variance and which shots were drawn. This is an argument about metric design, and it is the strongest thing this project has.

What was found

Finding	Value	Confidence
ψ target is a low-dimensional family	corpus median k99 = 5, median PR 1.222; batch 0 recomputed k99 = 7, PR 1.82499, top-3 0.94752	`[M]` corpus-wide
The oracle rank-7 floor and the gap to it	floor 0.0913 at k=7 (σ_meas); fitted ridge 0.838 → a 9.2x gap. The bound uses true coefficients, so it is not a target — it says the error lives in predicting ~7 coefficients, not in the representation	`[M]` batch 0
NRMSE is tail-dominated	top 1% of windows → 47.9% of test MSE (leaky split 17.5%)	`[M]`
NRMSE is variance-dominated	corr(RMSE, local σ) = 0.970; normalised spread 1.20x, U-shaped, worst at the edge	`[M]`
Error in derived quantities is EDGE-concentrated	flux-surface displacement 0.024 → 0.687 cells as t goes 0.1 → 0.9; core:edge 0.035	`[M]` pilot
Energy and geometry disagree in sign	median: elementwise energy 0.43–0.54× (better) vs level-set displacement 1.17–1.47× worse, both regions. Only the median phrasing is quotable	`[M]` pilot
EFIT frames entirely NaN	29.4% mean, 27.2% median (19.6–66.7%)	`[M]` corpus-wide
Dead input channels	1 of 115 dead in every batch; 3 more in 84.5%; rest batch-specific. No corpus-wide mask exists	`[M]` corpus-wide
Effective input dimension	batch-dependent: 4,415 (batch 0) vs 5,490 (12-batch subset)	`[M]`
`thomson_scattering` availability	90.11% mean (median 96.5%, min 19.5%)	`[M]` corpus-wide
ψ units	`Wb / rad`, in the zarr attrs but not TokaMark's stats file	`[M]`
Licence	card CC BY 4.0; data attrs CC BY-SA 4.0	`[M]`
Unscored signals present	`j_tor` (65,65,T); `vloop_dynamic`/`vloop_static` — unselected, not absent	`[M]`
Strongest baseline, Task 1-3	TokaMind FT-Base 0.122, 6,927,799 params, random split, 9,270 shots. CNN 0.1481; CNN+LSTM 0.1591. FT-Tiny ties 0.122	`[P]` from source
Classical baseline is per-task	CNN+LSTM beats CNN on 9 of 14; on 2-3 the bar is 0.1263, not 0.1358	`[P]`
Our in-house NRMSE is not protocol-comparable	dense/mean-imputed vs upstream's masked available-ground-truth; σ pinned to 0.03761 vs the YAML's 0.0549 — a 1.46–1.49x lever	`[M]`+`[P]`
Split matters more than model family	1-3 goes 0.1481 → 0.5016 (CNN), random → temporal	`[P]`

Tags: [M] measured · [R] repo · [P] paper · [I] inference (carries a falsifier) · [U] unverified.

Scope caveat. The corpus sweep is complete, so availability, dead-channel and manifold figures are corpus-wide. The region/estimator figures rest on a 122-shot pilot with an unreproducible split (X2, X35) and are labelled as such — they are [M] for that slice, never corpus-level claims. Batch 0 is the 1.7th percentile on Thomson availability and the 98.3rd on manifold width.

File	Role	Status
ROADMAP.md	Orientation; entry point for a new reader	✅ v2.3
GOVERNANCE.md	Rules, continuity plan, errata ledger + open questions (canonical), §4.5 identifier cross-index	✅ v2.1
DATASHEET.md	Code book, task contracts, descriptives, baselines	✅ v0.5
RESULTS.md	Results ledger `RES-1`–`RES-9` (question, n, result, counter-argument, falsifier)	✅ v1.3
ERRATA.md	Upstream-facing errata and data properties; Part C frozen with `→ X##` anchors	✅ v2.3
SPEC-P2.md	Grad–Shafranov-structured flux-map decoder	✅ v3.1 — passes the gate
SPEC-P1.md	Grey-box inductive response model (Group 2)	✅ v1.1
SPEC-DATASET.md	Windowed corpus artifact + metric suite	⚠ v1.0 — Stage-0 gated
DEV-ROADMAP.md	Gates, workstreams, cost ledger, risk register	✅ v2.2
FRICTION.md	Upstream report — drafted, not yet filed	✅ v1.1
HF-SCIENCE.md	Community ask to Hugging Science; contributions `HF-1`–`HF-3`	✅ v1.3
W1-MEMO.md	Interpretation memo — what the diagnostics license, and how to think about the data	✅ v1.0
HANDOFF.md	Recovery path and current state	✅ v4

No artifact fails the content gate. Verification debt and open gates are recorded in GOVERNANCE.md §7.

Evidence base

Measured quantities come from graziul/tokamark-p2-runs:

w1/recon.json — zarr structure, units, licence attrs, group availability
w1/summary.json — batch-0 missingness, manifold, ridge pilot, region-stratified error (partly superseded by X30/X39)
w1/singular_values.npy — the batch-0 target spectrum, float32, len 4225 — the source for X39/X40
w1/keep_idx.npy — batch-0-specific; must not be used as a corpus-wide mask
sweep/SUMMARY.json, sweep/batch_*.json — corpus-wide: availability, dead-channel classification, 58 per-batch records
q11/q11_derived_v4.json — flux-surface displacement (X33)
q12/q12_verify.json — σ, tail concentration, geometry, spectrum, truncation curve (X34, X36, X38, X39, X40)
q13/q13_verify.json — mechanism falsification (X45)
q15/q15_2x2.json — the region × estimator 2×2 (X48, X49)
gate/grep_gate.txt — the rule R14 verification run that found X51/X52
lcurve/X_*.npy, lcurve/Y_*.npy — 12 shards, 21% of the corpus, ~3.7 GB. X (193,346 × 5,515), Y (193,346 × 4,225). Caveats: the split (159,239/34,107; held-out shards 2 and 9) exists only in a job log; Y is present for the held-out shards; the windower SHA is unpinned (Q6); per-shard window counts spread 1.9x. lcurve_fits.json does not exist — the producing job was cancelled at 5% of windows