ProbeBench: classical reference selection

Measured focal-plane exposures from a simulated wavefront-control benchmark. Each record connects a detector image to the applied optical-path command, photon budget, episode identity, and exact signed detector array.

This release contains 30,912 exposures from 4,032 episodes, covering 14 classical candidates and 12 independent base wavefronts. It retains every candidate, including configurations that were not selected, and all optimizer diagnostics. The single split, public_tuning, supports inspection and reproduction of parameter selection. It does not provide held-out performance estimates.

Code · Methods and results · Native records · Measurement replay

Load an exposure

from datasets import load_dataset

frames = load_dataset(
    "noetherlabs/probebench",
    split="public_tuning",
    revision="classical-selection-v1",
    streaming=True,
)
frame = next(iter(frames))
frame["image"]  # PIL image; display rendering, not detector values

The images are embedded in Parquet shards. Streaming reads avoid creating a second complete dataset cache. The revision tag identifies this fixed release; record the resolved commit when reporting derived results.

Read the detector values

The raw_archive column identifies a ZIP shard. Its raw_array member contains the original NumPy file, including signed float64 values. Loading a member into memory requires no archive extraction.

import hashlib
import io
import zipfile

import numpy as np
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    "noetherlabs/probebench",
    filename=frame["raw_archive"],
    repo_type="dataset",
    revision="classical-selection-v1",
)
with zipfile.ZipFile(archive) as z:
    payload = z.read(frame["raw_array"])
assert hashlib.sha256(payload).hexdigest() == frame["raw_array_sha256"]
electrons = np.load(io.BytesIO(payload), allow_pickle=False)

Record structure

Fields	Meaning
image, image_sha256	Original PNG bytes and their SHA-256 digest.
episode_id, exposure_index, file_name	Identity and original image filename.
applied_opd_nm, expected_photons, cumulative_expected_photons	Applied command and acquired-exposure budgets.
candidate_id, family, selected_reference	Candidate identity and the global selection flag for each family.
base_wavefront_id, member, stratum_id, replicate	Dependence structure: base wavefront, conjugate member, RMS stratum, and detector replicate.
raw_archive, raw_array, raw_array_sha256	Archive, member path, and digest of the unchanged detector array.
evaluator_wfe_nm, evaluator_strehl_peak_ratio	Latent simulator labels for analysis.

Evaluator labels, cohort identities, selection flags, and simulator seeds are not controller observations. episodes.jsonl preserves complete native episode records, including optimizer diagnostics, seeds, latent wavefronts, and final-command observation indices. A final command measured earlier is distinct from the last exposure or an unmeasured submission.

selected-settings.json records six globally selected family configurations. configuration.json specifies the experiment. source-manifest.json binds the original ImageFolder export; source-metadata.jsonl and source-README.md preserve its metadata and format description. provenance.json records the native archive and rendering provenance. The publication manifest describes the Parquet and ZIP layout. These source artifacts contain evaluator information and are intended for reproduction and analysis.

Model and interpretation

The simulator uses static, monochromatic scalar Fourier optics with six controlled modes, a matched estimator, and Poisson–Gaussian detector noise. Conjugate members, detector replicates, and shared budget prefixes are correlated; the experiment has 12 independent base wavefronts, not 4,032 independent wavefronts.

PNGs display log1p(max(electrons, 0)) / log1p(photons_per_exposure) with fixed limits 0–1, the magma colormap, and lower image origin. The display clips negative values; the raw arrays preserve them. Every image corresponds to a retained exposure. No image is synthesized for an unmeasured command.

These data establish neither telescope performance nor language-model performance. Recorded decision times do not establish comparable algorithm latency or a real-time control deadline. Keep failed and nonselected candidates in analyses, and separate tuning results from any subsequent evaluation.

Citation

@misc{noetherlabs2026probebenchselection,
  title  = {ProbeBench: Classical Reference Selection},
  author = {{Noether Labs}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/noetherlabs/probebench}
}

Released under Apache-2.0. See LICENSE and NOTICE.