---
license: cc-by-4.0
language:
- en
- zh
pretty_name: SII Stellarator Configuration Dataset
tags:
- fusion
- stellarator
- quasi-isodynamic
- scientific-machine-learning
- tabular
configs:
- config_name: default
  data_files:
  - split: train
    path: main/*.parquet
- config_name: failed
  data_files:
  - split: train
    path: failed/*.parquet
- config_name: cons_bench
  data_files:
  - split: train
    path: cons_bench/*.parquet
---

# SII Stellarator Configuration Dataset

[中文](README.zh.md)

**Wenyang Li** | PhD student, [Shanghai Innovation Institute](http://www.sii.edu.cn)

[AI-Driven Controlled Fusion Simulation, Control & Design Lab](https://github.com/SII-AI4Fusion)

Email: [lwydsg@mail.nankai.edu.cn](mailto:lwydsg@mail.nankai.edu.cn) | Phone: [+86-156-2003-5216](tel:+86-156-2003-5216). License: CC BY 4.0.

## 2026-10-21: SOTA Configuration Update

**This release includes the configurations from our first-place ConStellaration MHD-stable QI submission and related CMA optimization evaluations.**

On the [official leaderboard](https://huggingface.co/spaces/proxima-fusion/constellaration-bench), the submission by **Wenyang_Li** achieved a hypervolume score of **248.227515058408**, ranking **1st** in the saved **2026-10-19** leaderboard snapshot. Submitted on 2026-10-18, it scored approximately **1.25%** above the runner-up's 245.1589221336.

All **783** submitted configurations have been verified by exact `configuration_id` lookup in the prepared release across `main/` and `cons_bench/`.

![ConStellaration MHD-stable QI leaderboard comparison](assets/cma_leaderboard.png)

Official leaderboard comparison from the 2026-10-19 snapshot.

![CMA Pareto fronts for NFP 2, 3 and 4](assets/cma_pareto.png)

CMA search results for NFP 2, 3 and 4, from the article's 2026-10-20 analysis snapshot. Stars mark reevaluated starting configurations; colored points show the evolved Pareto fronts in aspect ratio and L_delta_b.

## Overview

Stellarator boundary configurations and recorded physics evaluations, covering NFP 1-5.
This replaces the earlier 158,685-row release.
The original data came from [proxima-fusion/constellaration](https://huggingface.co/datasets/proxima-fusion/constellaration).
New sources include sBO/CMA, traditional AL, earlier BO, DP, ExLLM and FusionLeague evaluations.
Incremental maintenance also collects reviewed benchmark physics evaluations.

<!-- dataset-stats:start -->
Collection updated (UTC, per-file snapshots): **2026-10-21**; schema: **stellarator-1.1**.

| Directory | Configurations | Parquet shards |
| --- | ---: | ---: |
| `main/` | 3,801,307 | 77 |
| `failed/` | 133,745 | 3 |

**1,064,122** main rows have all nine metrics.
<!-- dataset-stats:end -->

The main table retains successful evaluations. Failed records are stored separately.

## Fields and Quality

Flat column names are retained. Missing values are **null**.

| Fields | Type and meaning |
| --- | --- |
| `configuration_id`, `schema_version` | Strings: canonical configuration SHA-256 and format version |
| `boundary.n_field_periods` | int64: NFP |
| `boundary.r_cos`, `boundary.z_sin` | Nested float64 lists: stellarator-symmetric Fourier boundary |
| `metrics.*` | Nine nullable float64 diagnostics listed below |
| `evaluation.success` | Nullable bool; null means unknown, not failure |
| `evaluation.error`, `.fidelity`, `.solver_version` | Nullable strings: original error and recorded evaluation settings |
| `evaluation.feasibility` | Nullable float64: source score, not a universal constraint definition |
| `source.name`, `.path`, `.record` | Strings: source family, relative file path, record locator |
| `quality.iota_representation` | `signed`, `absolute_only` or `unknown` |
| `quality.metric_count`, `.flags` | Count of finite metrics and list of quality/transformation flags |

The `.field` abbreviations in a row retain that row's prefix, e.g. `source.path`.

| Metric suffix after `metrics.` | Meaning |
| --- | --- |
| `aspect_ratio` | Aspect ratio |
| `edge_rotational_transform_over_n_field_periods` | Edge iota divided by NFP |
| `average_triangularity` | Average triangularity |
| `max_elongation` | Maximum elongation |
| `edge_magnetic_mirror_ratio` | Edge magnetic mirror ratio |
| `vacuum_well` | Vacuum magnetic well |
| `minimum_normalized_magnetic_gradient_scale_length` | L_delta_b |
| `qi` | QI diagnostic, not log10(QI) |
| `flux_compression_in_regions_of_bad_curvature` | Bad-curvature flux compression |

- **Deduplication:** exact NFP plus canonical Fourier coefficients. Prefer successful, then unknown, then failed evaluations; within a status prefer more complete metrics, then deterministic source order.
- **Boundary:** usually 5x9; higher-order matrices such as 8x15 are also retained. For `(M, 2N+1)`, rows are `m=0..M-1`, columns `n=-N..N`. Smaller supported grids are zero-padded; higher modes are preserved. Coefficients are not normalized or approximately rounded for deduplication.
- **Use limits:** metric completeness does not establish physical feasibility or comparable solver fidelity.

## Read the Data

```python
import pyarrow.dataset as ds
data = ds.dataset("main", format="parquet")
complete = data.to_table(filter=(ds.field("quality.metric_count") == 9)
                        & (ds.field("evaluation.success") == True))
```

For signed-iota training also require `quality.iota_representation == "signed"`;
for raw-QI-only studies exclude `qi_reconstructed_from_log10` in `quality.flags`.

## Maintain the Dataset

`FusionLeague/datasets/stellarator/` is the editing source: `main/`, `failed/`, these two READMEs and `tools/`.
Edit here first; the external HF clone is only a publication copy. Tools and audit archives are not uploaded.
**Step 1: update the source dataset on Linux.**

Start from the `FusionLeague/` repository root:

```bash
cd datasets/stellarator && python tools/update_dataset.py
```

If dependencies are missing, run `python -m pip install -r tools/requirements.txt` once, then retry.
Wait for a successful update before copying; do not update data while copying. Add `--dry-run` to inspect sources only.

- **Sources:** automatically scan `methods/` for real `qi`/`qi_frontier` evaluation journals; skip quadratic/examples and duplicate run exports. Collect `benchmarks/qi_design/data` submissions only when review, result hash, submission identity and point inputs agree. Running or unapproved results are deferred. No solvers or surrogate inference run.
- **Incremental merge:** index the current dataset on the first run, then remember completed log positions. Read new complete records, defer unfinished final lines, and apply the same exact deduplication/whole-record preference above. Unchanged sources are skipped; changed log prefixes are re-imported without deleting historical configurations. Revoked reviews do not automatically retract previously collected data.
- **Output:** without winning additions/replacements, leave Parquet and READMEs unchanged. Otherwise export and fully verify a sorted dataset, replace `main/` and `failed/`, and refresh only the statistics blocks in both READMEs. Ingestion is incremental, but changed datasets are fully re-exported and verified.
- **State and recovery:** `tools/.update-state/` holds the SQLite index, cursors, lock, latest summary, reports and previous published dataset. Only one updater may run. After an error/interruption, keep this state and rerun the same command to recover; do not manually remove its lock or edit/replace indexed Parquet files. Do not read/copy/push during installation or after an unfinished update.
- **Resources and audit:** first indexing takes longer; allow space for the index, staged output and previous dataset in addition to current data. The latest `last_update.json` points to the report/verification/quarantine; invalid records are also kept in the index `issues` table. Review notices and quality/provenance before publication. Earlier publication backups are cleaned after a successful update; this is not a full version archive.

The original full-source builder remains in `tools/build_dataset.py` for separate rebuilds; routine
updates do not rescan all historical optimization directories. Original snapshot audits remain at
`../_archive/stellarator-20260927/releases/20260923-full-v1/`. Local fixture tests:
`python -m unittest discover -s tools -p 'test_*.py'`.

## Publish to Hugging Face

Run in this order: **download from the official Hub into the external clone → merge maintenance additions → review and publish**. Commands use Bash and assume `FusionLeague/` and the external clone are side by side.

**One-time setup (skip if already cloned).** Install Git LFS and run from the `FusionLeague/` root; these commands leave you in that directory:

```bash
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/SII-AI4Fusion/SII-Stellarator-Configuration-Dataset ../SII-Stellarator-Configuration-Dataset &&
git -C ../SII-Stellarator-Configuration-Dataset lfs install --local
```

**Step 1: download complete data from the official Hub.** Start from the `FusionLeague/` root. If you just ran an update from the maintenance directory, use `cd ../..` to return to the root. The external clone should be on `main` with no uncommitted changes:

```bash
cd ../SII-Stellarator-Configuration-Dataset &&
GIT_LFS_SKIP_SMUDGE=1 git pull --ff-only https://huggingface.co/datasets/SII-AI4Fusion/SII-Stellarator-Configuration-Dataset main &&
git -c lfs.url=https://huggingface.co/datasets/SII-AI4Fusion/SII-Stellarator-Configuration-Dataset.git/info/lfs lfs pull --include="" --exclude="" &&
git lfs fsck HEAD
```

This updates the Git revision, downloads all LFS files for that revision, then verifies their integrity. Continue to Step 2 after `Git LFS fsck OK`; you are now in the external clone. **For download-only use, stop here.** After an interrupted download, remain in the external clone and repeat the `git -c ... lfs pull ...` and `git lfs fsck HEAD` commands above; synchronize only after they succeed.

To use the mirror, replace the domain in both URLs with `hf-mirror.com`; the mirror may lag behind.

**Step 2: synchronize local additions into the external clone.** After preparing local maintenance data and completing Step 1, run the next commands without changing directories:

```bash
python -m pip install -r ../FusionLeague/datasets/stellarator/tools/requirements.txt &&
python ../FusionLeague/datasets/stellarator/tools/prepare_publish.py --hf-repo . &&
cp -r ../FusionLeague/datasets/stellarator/assets . &&
git add README.md README.zh.md assets &&
git diff --cached --stat
```

Install dependencies (including `duckdb`) into the current Python environment before running synchronization; the merge tool does not download from the Hub again. It uses the data downloaded in Step 1 as the baseline and merges maintenance data by `configuration_id`. Duplicate configurations keep a whole evaluation, preferring success, metric completeness and source order. `cons_bench` also merges submission memberships. Existing other subsets, extra columns, descriptions and configurations are retained. Output is installed and staged in the external clone; local maintenance data and its incremental index stay unchanged. The old delete-and-copy commands are no longer used.

The terminal shows merge, shard and verification progress. Incomplete downloads or uncommitted edits stop preparation. Temporary files and recovery state are kept in the external clone's `.git/stellarator-publish/`. Do not edit either dataset directory during merging.

**Step 3: review and publish.** Review the changes, then remain in the external clone and run; skip if nothing changed:

```bash
git commit -m "Update consolidated stellarator dataset" &&
git lfs fsck HEAD &&
git push origin main
```

Commands use Bash; stop on any error. Commit, verify LFS, then push; tools, caches and archives are not uploaded. If push is rejected, do not force-push: retain local commits, resolve Git divergence, then synchronize and merge again. Keep recovery state after an interrupted merge and rerun Step 2.
## Update Notes

This maintenance update (2026-10-17):

- Adds samples produced by an earlier combined optimization algorithm and merges them with existing data using the established deduplication rules.
- Adds the separate `cons_bench` subset, collecting public configurations and published evaluations submitted by ConStellaration leaderboard users while retaining submission provenance. Thank you to the participants for sharing their results; source and licensing details are provided below.

The dataset will receive incremental updates from time to time. Follow [SII-AI4Fusion on Hugging Face](https://huggingface.co/SII-AI4Fusion) and like [this dataset](https://huggingface.co/datasets/SII-AI4Fusion/SII-Stellarator-Configuration-Dataset) to stay connected with future updates.

## ConStellaration Leaderboard Solutions

`cons_bench` is a separate subset of public results for all three tasks on the [ConStellaration leaderboard](https://huggingface.co/spaces/proxima-fusion/constellaration-bench). It is stored in `cons_bench/` and is not merged into `main/`. Run from this dataset directory:

```bash
python -m pip install -r tools/requirements.txt
python tools/update_cons_bench.py
```

The source is the leaderboard's [public results repository](https://huggingface.co/datasets/proxima-fusion/constellaration-bench-results). Repeated runs download only new or changed files. A successful run updates the separate Parquet and bilingual statistics and registers the HF `cons_bench` subset with a `train` split. Keep `tools/.cons-bench-state/` after a network interruption and rerun the same command. To use the mirror, run `HF_ENDPOINT=https://hf-mirror.com python tools/update_cons_bench.py`.

Each exact unique configuration has one row. `benchmark.entries` retains all submission links, public usernames, times, tasks, scores, per-task snapshot ranks and point objectives/feasibility, preserving repeated submission provenance. MHD scores belong to whole submitted sets, not individual points. `benchmark.boundary_json` retains the representative's original boundary; `benchmark.problem_types` lists its tasks. Row-level metrics come from the first entry in `benchmark.entries`, preferring the evaluation with more available metrics.

Only published metrics are retained: maximum elongation for geometrical, L_delta_b for simple-to-build, and aspect plus L_delta_b for MHD. Unpublished metrics, solver status and versions remain null; VMEC is not rerun. `source.revision` and pinned source links identify the upstream snapshot. `source.license` retains the upstream license ID or null when unspecified; this repository's CC BY 4.0 declaration does not grant rights on behalf of the upstream source. Shared `configuration_id` values can overlap the main table, so this subset is not automatically a disjoint test set. Submissions by user `test` are excluded, matching the leaderboard page.

<!-- cons-bench-stats:start -->
Publication merge: **2026-10-21**; 163 submissions, 15,619 unique configurations.
<!-- cons-bench-stats:end -->

```python
from datasets import load_dataset
bench = load_dataset("SII-AI4Fusion/SII-Stellarator-Configuration-Dataset",
                     "cons_bench", split="train")
```

## 2026-10-17 CMA / NGO Incremental Update

- CMA: this round added **37,632** configurations to the maintenance source and replaced **1** existing record.
- NGO: this round added **101,190** configurations to the maintenance source.

In total, **138,822** configurations were added and **22** duplicate records were ignored; the export passed verification. These are additions relative to the local maintenance source, not net additions to HF. Publication retains existing HF data and recomputes the final total after merging and deduplication.