---
language:
- en
- zh
configs:
- config_name: default
  data_files:
  - split: train
    path: main/*.parquet
- config_name: failed
  data_files:
  - split: train
    path: failed/*.parquet
---

# SII Stellarator Configuration Dataset: Consolidated Format

Schema version: `stellarator-1.1`. This card describes the new offline export,
not a claim that the existing Hugging Face dataset has already been replaced.
See `report.json` for this particular export's actual counts and `inventory.json`
for its source file inventory. Upload and source-license review are separate steps.

## One Row, One Configuration

The primary data is `main/*.parquet`. It includes successful evaluations even when
physical constraints are violated, plus explicitly labelled unknown-status partial
evaluations. `failed/*.parquet` contains explicit solver failures with identifiable
boundaries. These are not train/test splits. The HF `train` label above is only a
loader convention; no evaluation holdout has been defined.

Identical NFP and numerical Fourier coefficients produce one record across both
groups. Prefer successful over unknown over failed status, then more finite
metrics, then deterministic source ordering. Repeated evaluations are not kept as
separate samples. A selected row is kept whole; missing fields are never filled
from a different evaluation of the same boundary.

## Boundary

| Column | Type | Meaning |
| --- | --- | --- |
| `configuration_id` | string | SHA-256 of canonical NFP, mode grid and float64 coefficients |
| `schema_version` | string | `stellarator-1.1` |
| `boundary.n_field_periods` | int64 | Positive integer NFP; different NFP means different configuration |
| `boundary.r_cos` | list[list[float64]] | Radial cosine coefficients; usually 5 by 9 |
| `boundary.z_sin` | list[list[float64]] | Vertical sine coefficients; same shape as r_cos |

For shape `(M, 2N+1)`, rows are poloidal modes `m=0..M-1` and columns
are toroidal indices `n=-N..N`. Standard records use 5 by 9; DP also uses 8 by 15.
`r_cos[0, n<0]` and `z_sin[0, n<=0]` must be zero under the source convention.
Only stellarator-symmetric boundaries are represented; nonzero `r_sin` or `z_cos`
are rejected, not discarded silently. Coefficients are never rescaled by major
radius or standardized as model features. Recognized smaller centered odd-width
arrays are padded to at least 5 by 9; larger arrays retain every coefficient,
with a `higher_fourier_modes_preserved` flag. Flat length-45 arrays are reshaped row-major. Negative zero becomes
zero. All other float64 values are retained exactly, with no tolerance/rounding.

For the standard 5-by-9 grid, the ID hashes little-endian signed int64 NFP followed by float64 values in row-major
`r_cos`, then row-major `z_sin` order (`struct.pack('<q90d', ...)`). This also makes
JSON formatting and equivalent zero padding up to the standard grid irrelevant to deduplication.
For a larger grid, the prefix is `b'higher-modes\x00'`, followed by NFP/int64,
row count/int32 and column count/int32, then all coefficients in the same order.
No higher coefficient is rounded or discarded to match the standard grid.

## Nine Physical Metrics

Every field is nullable float64. Null means unavailable or nonfinite, not zero.

| Column | Meaning / legacy pool name |
| --- | --- |
| `metrics.aspect_ratio` | Aspect ratio / `aspect_ratio` |
| `metrics.edge_rotational_transform_over_n_field_periods` | Edge iota divided by NFP / `iota_nfp` |
| `metrics.average_triangularity` | Mean triangularity / `triangularity` |
| `metrics.max_elongation` | Maximum elongation / `elongation` |
| `metrics.edge_magnetic_mirror_ratio` | Edge magnetic mirror ratio / `mirror` |
| `metrics.vacuum_well` | Vacuum well / `vac_well` |
| `metrics.minimum_normalized_magnetic_gradient_scale_length` | L_delta_b / `l_delta_b` |
| `metrics.qi` | QI diagnostic, not log10(QI) |
| `metrics.flux_compression_in_regions_of_bad_curvature` | Bad-curvature flux compression / `flux_comp` |

Source metric conventions are retained. Signed iota is not turned into an absolute
value. Historical BO data that already lost its sign remains positive/absolute-only
and is labelled accordingly. Its saved `qi_log10` is converted using `10**value`;
this cannot undo historical clipping and is marked `qi_reconstructed_from_log10`.
No predicted metric, current constraint score or newly computed physics value is
substituted for a missing original diagnostic.

## Evaluation, Source and Quality

| Column | Type / interpretation |
| --- | --- |
| `evaluation.success` | Nullable bool; null means original status cannot be established |
| `evaluation.error` | Nullable string; original error or explicit original forward-model failure flag |
| `evaluation.fidelity` | Nullable string; recorded fidelity only |
| `evaluation.solver_version` | Nullable string; recorded execution version, otherwise recorded configuration version |
| `evaluation.feasibility` | Nullable float64; source score, not a unified feasibility definition |
| `source.name` | Source family selected by the importer |
| `source.path` | Relative source file path, without local machine root |
| `source.record` | Parquet zero-based row; JSONL one-based line; nested objects add path components |
| `quality.iota_representation` | `signed`, `absolute_only` or `unknown` |
| `quality.metric_count` | Number of finite values among the nine metrics |
| `quality.flags` | List of transformation, invalid-value or status-inference labels |

Missing solver versions, fidelity and timestamps are not inferred from today's
code. Pool success may be inferred from its documented source filtering; the flag
`success_inferred_from_source` distinguishes this from an explicit result status.
Complete unlabelled historical measurements can have
`success_inferred_from_complete_metrics`. A partial record with no reliable status
uses `evaluation_status_unknown`. Invalid scalar metrics carry
`nonfinite_or_invalid:<metric>`. Unknown iota convention must not be treated as proof
that its sign was preserved. Metric completeness does not establish fidelity or
physical feasibility.

`higher_fourier_modes_preserved` requires downstream models to support variable
mode grids. `legacy_boundary_text_precision` identifies ExLLM arrays stored as
rounded printed numbers; missing digits cannot be reconstructed. DP raw constraint
values are retained, but its scaled objective is not assumed to equal raw L_delta_b
(`scaled_target_not_used_as_raw_ldb`).

## Loading and Filtering

```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))
# Request just the original 12 columns for compatible downstream consumers.
core = complete.select([name for name in complete.column_names
                        if name.startswith(('boundary.', 'metrics.'))])
```

For signed-iota training, additionally require `quality.iota_representation ==
'signed'`. For raw-QI-only studies, exclude records carrying the reconstruction
flag. `failed` is for failure analysis, not a source of zero-valued physics labels.
No common `feasibility <= threshold` filter is guaranteed valid across all sources.

## Coverage and Review

The builder recognizes cleaned/raw original Parquet, historical and current
optimization pools/CMA/sBO records, legacy BO, traditional AL final results, DP
SQLite/NPZ evaluations, ExLLM stored measurements and
FusionLeague QI/QI-frontier evaluation records. It excludes leaderboard references,
quadratic tests, surrogate predictions, latent-only points and opaque binary model
or optimizer state. Unsupported formats are not interpreted as physical labels.

Input files are individually bounded snapshots, not one global freeze. Malformed
or unusable records are indexed in `quarantine.jsonl` without fabricating a row;
all source files should be retained for investigation. `report.json` lists duplicate
counts, missing data, selected source distribution, read statistics and shard
checksums. Check those reports before publication. Scientific comparability and
licenses across sources require review; this export does not grant new permissions.