---
language:
- en
license: apache-2.0
pretty_name: ProbeBench classical reference selection v1
tags:
- optics
- wavefront-sensing
- scientific-computing
configs:
- config_name: default
  data_files:
  - split: public_tuning
    path: images/**
---

# ProbeBench classical reference selection v1

This export retains 4032 episodes and 30912 acquired images across all
14 candidates and 12 independent base wavefronts. The split is
`public_tuning`: these are public parameter-selection data, not training/test splits,
held-out performance estimates, telescope measurements or language-model results.

`images/metadata.jsonl` preserves candidate, family, photon/exposure cell, base,
conjugate member, RMS stratum and detector-replicate identity. `selected_reference`
identifies the single global choice per family; non-winning candidates remain included.
The command and expected-photon fields describe the actual exposure. Columns prefixed
`evaluator_` are latent simulator labels and must not be supplied as controller inputs.
Cohort identity and selection flags are analysis labels, not additional observations.

`raw/` stores the corresponding signed float64 electron arrays. `raw_array` paths are
relative to this export root, while `file_name` is relative to `images/`. PNGs use
log1p(max(electrons, 0))/log1p(photons per exposure), fixed limits 0–1, the magma map,
and image origin lower. Display clipping does not alter the raw arrays. No final image
is invented for an unmeasured command; every image corresponds to a retained exposure.

`episodes.jsonl` preserves complete native case records under `record`, including
latent wavefronts, seeds, all optimizer diagnostics, and final-command observation
indices. A final command measured earlier is distinct from the last exposure and from
an unmeasured submission. Episode records and `configuration.json` contain evaluator
information and are not observation inputs. `selected-settings.json` preserves the
six globally selected configurations. `provenance.json` binds the source fingerprints,
full replay receipt, generator and rendering dependencies. `manifest.json` inventories
this export separately from the original archive.

## Load locally without a second image cache

```python
from pathlib import Path
from datasets import load_dataset
import numpy as np

root = Path("work/selection-imagefolder")
frames = load_dataset("imagefolder",
    data_files={"public_tuning": str(root / "images" / "**")},
    split="public_tuning", streaming=True)
first = next(iter(frames))
electrons = np.load(root / first["raw_array"], allow_pickle=False)
```

The layout follows the official [ImageFolder metadata format](https://huggingface.co/docs/datasets/image_dataset)
and [local loading configuration](https://huggingface.co/docs/datasets/loading).

The model is static monochromatic scalar Fourier optics with six controlled modes,
an exactly matched estimator, and Poisson–Gaussian detector noise. Conjugates,
replicates and shared budget prefixes are correlated; they do not add independent
wavefronts. Keep all numerical failures and non-winning results in analyses. Recorded
decision times do not establish comparable algorithm latency or real-time guarantees.
Hashes establish consistency with the pinned release, not authorship. Shared file
locking excludes cooperating writers; keep the input tree unchanged during export.
An interrupted export has no completed manifest and must not be treated as a dataset.