BESS-Bench: Long-baseline, mixed-quality Be-star spectra for ML time-series research

339 115 spectral rows · 37 624 physical observations · 1 468 Be stars · 35 years · 91.6 % of observations from amateur observers A benchmark for representation learning and temporal modelling on heterogeneous stellar spectroscopic time series.

Three granularity levels (see §Dataset composition): parquet rows (atom of the file), physical observations (one observing session ≡ one (star_name, mjd) group; an échelle observation spans a median of 26 orders), and unique stars (split unit). The amateur share is $91.6,%$ at the observation level and $81.7,%$ at the row level (échelle orders multiply professional row counts).

TL;DR

BESS-Bench is an ML-ready, per-star split release of the BeSS community Be-star spectral archive, together with:

• Full-range optical spectra (1 150 – 10 442 Å native coverage). The fraction of spectra that strictly cover each commonly studied Be-star line at ±50 Å is reported at two granularities — per physical observation (Hα 77.6 %, Hβ 35.7 %, He I 5876 36.0 %, Na D 36.0 %, Hγ 32.5 %, Fe II 5169 35.6 %, Hδ 19.4 %, O I 7772 4.0 %) and per parquet row (row-level coverage used inside tasks: Hα 7.9 %, Hβ 0.8 %, …). Both tables are reproduced in §Line coverage below.
• A pre-trained Hα-window reference encoder, BeMAE (≈795 K parameters, three seeds), released separately as matthieulel/bemae-halpha-v1.
• Three benchmarked downstream tasks (SpecProbe feature regression, LineTransfer Hβ → Hα generalisation, EWForecast short-horizon EW(Hα) forecasting), with fixed protocols and bootstrap / multi-seed confidence intervals. The three tasks focus on Hα and Hβ because these are the only two lines with row-level coverage large enough to support meaningful confidence intervals in v1.0.
• A task-dependent ranking: the learned encoder dominates every linear representation on line-shape features (SpecProbe), but does not beat persistence on EWForecast, where a simple PCA+Ridge baseline does (MAE ratio 0.942, bootstrap CI95 [0.927, 0.957]) and is ahead of nine zero-shot time-series foundation-model checkpoints; a linear head trained on top of frozen Chronos-Bolt-base reaches 0.932.

Full methodology, datasheet and reproducibility instructions are in the accompanying paper and in DATASHEET.md. Result-level JSON artefacts (per-task summaries, bootstrap CIs, coverage audits) live in the companion code repository; this card quotes their content for self-containedness.

Quick start

from datasets import load_dataset

# Full dataset (10 GB, 21 shards)
ds = load_dataset("matthieulel/bess-bench", split="train")

# Streaming access (no full download)
ds_stream = load_dataset("matthieulel/bess-bench", split="train", streaming=True)

# A small stratified sample (1,976 spectra, 150 stars, ~70 MB) for quick inspection
sample = load_dataset("matthieulel/bess-bench-sample", split="train")

# Apply the official per-star split
import pandas as pd
splits = pd.read_csv("hf://datasets/matthieulel/bess-bench/splits.csv")
train_stars = set(splits.query("split == 'train'").star_name)
train = ds.filter(lambda x: x["star_name"] in train_stars, num_proc=8)

Reproduction of every paper table from frozen checkpoints is documented in REPRODUCE.md of the companion code repository.

Dataset composition

Three granularity levels

BESS-Bench is distributed as a parquet table, and we distinguish three nested units of measurement:

Unit	Count	Definition
Parquet rows	339 115	One row = one 1D spectrum or one échelle order
Physical observations	37 624	One observing session ≡ one (star_name, mjd)
Unique stars (split unit)	1 468	Per-star split, no leakage

The ~9× inflation from physical observations to parquet rows comes from 9 917 échelle sessions each stored as 19–40 individual orders (median 26 orders / session) — an intentional choice that preserves the native resolution (R ≈ 10 000 – 17 000). The 27 707 remaining observations are single-range spectra (one row = one observation).

Stream	Rows	Physical obs.	Unique stars	Median R
Single-range	62 476	27 707	1 354	9 000
Échelle orders	276 639	9 917	915	11 000
Total	339 115	37 624	1 468	—

Summary metrics

Metric	Value
Parquet rows	339 115
Physical observations	37 624
Unique stars	1 468
Temporal coverage	1990-03 → 2025-12 (35 years)
Wavelength coverage	1 150 – 10 442 Å
Median SNR	189
Amateur fraction (rows)	81.7 %
Amateur fraction (obs.)	91.6 %
Spectrographs / telescopes	26 / 62
Licence	CC-BY-4.0

Filter échelle orders with the is_echelle_order flag. v1.0 downstream tasks group by (star_name, mjd, instrument_setup) so that an échelle observation contributes once, not 26 times.

Wavelength frame convention

The wavelength column is shipped as in the BeSS FITS (no shift applied), so users can reproduce any rest frame they wish. For ≈ 97 % of spectra the released grid is therefore topocentric. To reach the heliocentric frame, apply

λ_helio = λ_obs · (1 − bss_rqvh / c)         (c = 299 792.458 km/s)

helio_velocity (= FITS BSS_VHEL) is the correction already applied by the observer (0 for 97 % of spectra) and bss_rqvh (= FITS BSS_RQVH) is the residual correction still to apply (populated for 99.8 % of spectra, ±20 km s⁻¹, σ ≈ 10 km s⁻¹).

The bundled BESS-Bench baselines (encoder pre-training, SpecProbe/LineTransfer/EWForecast) apply this correction internally before the 128-bin Hα/Hβ crop so that features and embeddings live in a common heliocentric frame compatible with external surveys (SDSS, LAMOST, APOGEE).

Line coverage — two complementary tables

BESS-Bench ships the native wavelength coverage of each observation (not a cropped Hα window). Two coverage tables are computed on the v1.0 snapshot (audit script and full JSON in the companion code repository).

Table A — coverage per physical observation (astrophysical view: of the 37 624 observing sessions, which contain this line at ±50 Å? Échelle rows are aggregated via λ_min=min, λ_max=max over their orders.)

Line	λ₀ (Å)	Coverage	Observations covered (out of 37 624)
Hα	6562.8	77.63 %	29 207
Na D	5893.0	35.98 %	13 536
He I 5876	5876.0	35.97 %	13 532
Hβ	4861.3	35.66 %	13 416
Fe II 5169	5169.0	35.55 %	13 376
Hγ	4340.5	32.54 %	12 242
Hδ	4101.7	19.45 %	7 317
O I 7772	7772.0	4.05 %	1 523

Table B — coverage per parquet row (per-row coverage, used when tasks iterate over individual parquet rows rather than over observing sessions)

Line	λ₀ (Å)	Coverage	Rows covered (out of 339 115)
Hα	6562.8	7.94 %	26 937
Na D	5893.0	3.37 %	11 413
He I 5876	5876.0	1.94 %	6 570
Hβ	4861.3	0.78 %	2 652
Fe II 5169	5169.0	0.71 %	2 421
Hγ	4340.5	0.69 %	2 337
Hδ	4101.7	0.66 %	2 246
O I 7772	7772.0	0.41 %	1 387

The 10× gap between the two tables is explained entirely by the échelle decomposition: each échelle row covers a narrow sub-range by construction, so a single observing session that covers Hα contributes 26 rows in Table B but only 1 observation in Table A. Wavelength distribution across rows: λ_min p5/p50/p95 = 3 980 / 5 277 / 7 006 Å; λ_max p5/p50/p95 = 4 113 / 5 413 / 7 351 Å.

v1.0 baselined tasks (SpecProbe, LineTransfer, EWForecast) use the Hα and Hβ windows because they combine high observation-level coverage (77.6 % / 35.7 %) with enough row-level density (26 937 / 2 652 spectra) for stable CVs. He I, Na D and Fe II 5169 each reach ~36 % observation-level coverage and are released for community experimentation, but no v1.0 leaderboard task targets them.

Splits

Split by star (not by spectrum) to prevent temporal leakage:

Split	Stars	Spectra	Fraction
train	1 024	241 057	≈71 %
validation	219	44 383	≈13 %
test	225	53 675	≈16 %

Assignment is deterministic via md5("bess_bench_v1::" + star_name)[:8] mod 100 ↦ {0–69: train, 70–84: val, 85–99: test}. The result is bundled as splits.csv; the generating script is shipped in the companion code repository for full reproducibility.

How the splits are used. This star-level partition is the reference split for submissions that train on labels. The v1.0 baselines of the paper do not need it: SpecProbe and LineTransfer use 5-fold GroupKFold on star_name over all scored stars, EWForecast uses a per-star temporal cutoff (MJD 59215), and the BeMAE encoder is pre-trained without labels on all Hα spectra except those of eight held-out stars (GAM CAS, PLEIONE, ZET TAU, V442 AND, PI AQR, DEL SCO, 28 CYG, OME CMA).

Schema

Column	Type	Description
wavelength	list[float32]	Native wavelength grid (Å; in nm for 2 311 rows, see Known issues)
flux	list[float32]	Calibrated flux
flux_error	list[float32]?	Flux uncertainty (null for a substantial fraction of amateur FITS; see DATASHEET §3)
spectrum_id	string	MD5 (star_name, mjd, instrument_setup)
star_name	string	Canonical de-duplicated name (1 468 unique)
star_name_raw	string	Original BeSS object name
ra, dec	float64	J2000 (degrees)
lambda_min, lambda_max	float64	Spectral coverage (Å)
n_pixels	int32	Number of spectral pixels
spectral_resolution	float64	Theoretical R
spectral_resolution_measured	float64?	Measured R (null when unavailable)
snr	float64?	DER_SNR (null for 29 corrupted spectra)
snr_continuum	float64	Continuum-window SNR
snr_quality	string	excellent/good/medium/low/very_low/unknown
observation_date	string	ISO-8601 date
mjd	float64	Modified Julian Date
exposure_time	float64	Seconds
temporal_quality	string	good / no_exptime / very_long
spectrograph	string	Normalised
telescope	string	Normalised
detector	string	Normalised
instrument_setup	string	Canonical (spectrograph, telescope) pair
instrument_confidence	string	high / medium / low
observer_type	string	amateur / professional / unknown
site_latitude	float64?	Degrees, rounded to 0.1° (≈11 km)
site_longitude	float64?	Degrees, rounded to 0.1° (≈11 km)
site_elevation	float64?	Metres, rounded to 100 m
helio_velocity	float64?	Heliocentric correction already applied by the observer (= FITS BSS_VHEL, km s⁻¹). 0 for ≈97 % of spectra.
bss_rqvh	float64?	Residual heliocentric correction to apply to reach the heliocentric frame (= FITS BSS_RQVH, km s⁻¹). ±20 km s⁻¹, 1σ≈10 km s⁻¹, populated for 99.8 % of spectra.
telluric_corrected	string	Raw FITS header string (free-form, often None)
normalized	string	Raw FITS header string (free-form, often None)
is_echelle_order	bool	True when the row is one order of an echelle
fits_format	string	Original FITS flavour

Benchmark protocol

BESS-Bench v1.0 releases three quantitative tasks (SpecProbe, LineTransfer, EWForecast). Every task reports reproducible mean ± std and/or bootstrap CI95 over seeds {42, 123, 456} or 10 000 bootstrap resamples, under the rule that a learned-model claim of improvement requires the corresponding CI to exclude the baseline value.

SpecProbe — Hα single-line spectral-feature regression (Ridge probe, 3 encoder seeds)

Frozen 128-D embeddings from the released Hα-window encoder → Ridge regression on six Hα spectral features. 5-fold GroupKFold on star_name (no star on both sides of a fold) over the 26 858 scored Hα spectra of 1 428 stars (Δv / vr_ratio are defined on the 10 804 cleanly double-peaked profiles), repeated for encoder seeds {42, 123, 456}. Baselines: the same probe on PCA(10/32/50) and on the raw 128-bin normalised flux. R², mean ± std over the three seeds (the linear baselines are deterministic):

feature	z (128-D, BeMAE)	PCA(10)	PCA(50)	raw flux (128-D)
ew	0.993 ± 0.002	1.000	1.000	1.000
peak_intensity	0.993 ± 0.003	0.950	0.947	0.948
fwhm	0.508 ± 0.037	0.094	0.125	0.121
central_depth	0.879 ± 0.023	-0.001	-0.028	-0.016
delta_v	0.412 ± 0.056	0.139	0.192	0.188
vr_ratio	0.129 ± 0.068	0.117	0.030	0.063

Aggregated ratio of mean R² over fwhm, central_depth, delta_v, z vs PCA(10): 7.77 (seed-level bootstrap CI95 [7.26, 8.06]). The encoder adds clear value on the non-linear shape features; EW and peak intensity are linearly decodable from the flux, and vr_ratio sits near its noise floor for every representation.

These numbers supersede those of the submitted manuscript, which had been computed with spectrum-level folds (stellar identity leaking across folds). The ranking is unchanged; absolute R² values are lower.

LineTransfer — Cross-line probing (Hβ → Hα features)

Same Hα targets as SpecProbe; the input is the Hβ ± 50 Å window (128 bins, locally normalised), restricted to the 2,525 same-night (star, MJD) pairs from 1,090 stars where both lines are observed. Ridge probe on a linear representation of Hβ, 5-fold GroupKFold on star_name; R², mean ± std across folds:

Hα target	pairs	Hβ PCA(10)	Hβ PCA(50)	Hβ raw flux
ew	2,525	0.66 ± 0.06	0.66 ± 0.06	0.71 ± 0.07
peak_intensity	2,525	0.28 ± 0.19	0.38 ± 0.10	0.17 ± 0.56
fwhm	2,508	0.00 ± 0.09	0.05 ± 0.02	0.00 ± 0.04
central_depth	2,525	-0.01 ± 0.07	0.03 ± 0.07	0.00 ± 0.04
delta_v	190	-0.08 ± 0.22	-0.11 ± 0.15	0.16 ± 0.14
vr_ratio	154	-0.04 ± 0.03	-0.02 ± 0.03	-0.07 ± 0.05

Equivalent width transfers (R² ≈ 0.66–0.71); line-shape parameters do not. delta_v and vr_ratio are defined on fewer than 200 pairs and are inconclusive. BeMAE is an Hα-only encoder and is not evaluated here.

EWForecast — Short-horizon EW(Hα) forecasting (bootstrap CI95)

One-step-ahead forecast of EW(Hα) at the next observation epoch given the past epochs of the same star. Primary metric: ratio_mae = MAE(model) / MAE(persistence) on the same test pairs. Split: per-star temporal cutoff at MJD = 59215 (2021-01-01); the PCA basis and the ridge heads are fitted on pre-cutoff epochs only. A model beats persistence when the bootstrap CI95 of the ratio (10 000 resamples of the test pairs) lies strictly below 1.

Test populations: 7,313 pairs from 621 stars for ctx=5 (6 664 pairs from 432 stars for ctx=10); zero-shot forecasters need no minimum context and are scored on 8 118 pairs from 846 stars. The elapsed time to the target is not a model input (next-observation forecasting).

Method	MAE ratio	CI95	beats persistence?
PCA(10)+Ridge, ctx=5	0.942	[0.927, 0.957]	yes
PCA(10)+Ridge, ctx=10	0.956	[0.937, 0.976]	yes
PCA(50)+Ridge, ctx=5	0.983	[0.966, 1.002]	inconclusive
PCA(50)+Ridge, ctx=10	1.040	[1.016, 1.065]	no
BeMAE z(128)+Ridge, ctx=5	1.003	[0.982, 1.026]	inconclusive
BeMAE z(128)+Ridge, ctx=10	1.014	[0.991, 1.038]	inconclusive
Chronos-Bolt-small (48 M), zero-shot	0.981	[0.966, 0.997]	yes
Chronos-Bolt-base (205 M), zero-shot	0.968	[0.955, 0.980]	yes
TimesFM-2.0 (500 M), zero-shot	0.969	[0.949, 0.989]	yes
Moirai-1.1-R-base (91 M), zero-shot	1.021	[1.002, 1.041]	no
Moirai-1.1-R-large (311 M), zero-shot	0.992	[0.975, 1.008]	inconclusive
Moirai-MoE-1.0-R-small (117 M), zero-shot	1.004	[0.985, 1.024]	inconclusive
Moirai-MoE-1.0-R-base (935 M), zero-shot	1.061	[1.034, 1.088]	no
Time-MoE-50M (113 M), zero-shot	0.988	[0.970, 1.006]	inconclusive
Time-MoE-200M (453 M), zero-shot	0.986	[0.968, 1.004]	inconclusive
Chronos-Bolt-base + trained linear head (ctx=5 pairs)	0.932	[0.918, 0.947]	yes
— control: ridge on the 5-epoch EW context	0.945	[0.931, 0.959]	yes
— Chronos-Bolt-base zero-shot on the same pairs	0.955	[0.941, 0.970]	yes
Causal Transformer on BeMAE embeddings (3 seeds, own harness)	1.17 ± 0.05	—	no

Zero-shot numbers are lower bounds on what trained or fine-tuned foundation models could reach. Median per-spectrum EW measurement uncertainty (Vollmann & Eversberg 2006 applied to the shipped DER_SNR): ≈ 0.44 Å over the stars with ≥ 30 epochs (0.46 Å over the scored slice), against a persistence MAE of 1.47 Å and a best MAE reduction of ≈ 0.09 Å: no current baseline is scientifically decisive yet.

Known issues in v1.0

The v1.0 data files are kept byte-identical to the version that was peer-reviewed. Two issues were found afterwards and will be fixed in the next data version:

• Wavelength unit. For 2 311 rows (0.7 %: 2 240 MUSICOS orders, 69 VLT/FLAMES spectra, 2 others) the wavelength array is stored in nm instead of Å, while lambda_min / lambda_max are in Å; one further row has inconsistent wavelength metadata. The affected spectrum_ids are listed in known_issues_v1.0.csv. Work-around: if lambda_max / max(wavelength) > 5: wavelength *= 10. None of these rows enters a scored target.
• Observer type. The 69 VLT/FLAMES rows (NGC 2004 members) are labelled observer_type = "amateur"; they are professional observations.

Citation

@inproceedings{bess_bench_2026,
  title     = {BESS-Bench: Benchmarking Spectral Representations for Be-Star Variability},
  author    = {Le Lain, Matthieu and Lefèvre, Sébastien},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://huggingface.co/datasets/matthieulel/bess-bench}
}

The frozen v1.0 snapshot of the data files is archived on Zenodo under the DOI 10.5281/zenodo.23808620 (the files are byte-identical to this repository). To cite the data themselves:

@dataset{bess_bench_data_v1,
  title     = {BESS-Bench v1.0: Be-star H-alpha spectra from the BeSS database},
  author    = {Le Lain, Matthieu and Lefèvre, Sébastien},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.0},
  doi       = {10.5281/zenodo.23808620},
  url       = {https://doi.org/10.5281/zenodo.23808620}
}

When using BESS-Bench, please also cite the underlying BeSS database:

@article{neiner2011bess,
  title   = {The BeSS database},
  author  = {Neiner, C. and others},
  journal = {Astronomical Journal},
  volume  = {142},
  pages   = {149},
  year    = {2011},
  doi     = {10.1088/0004-6256/142/5/149}
}

Licence

CC-BY-4.0. Attribution must credit the BeSS database (Neiner et al. 2011) and the BeSS consortium (LIRA / Observatoire de Paris). Commercial use is permitted under the same attribution conditions.

Acknowledgements

We thank the BeSS community of amateur and professional observers for 35 years of sustained spectroscopic monitoring. BESS-Bench packages and structures these public observations for ML use, and credits the original observers via the BeSS database.

Publications that use BESS-Bench should carry the acknowledgement requested by BeSS: "This work has made use of the BeSS database, operated at LIRA, Observatoire de Paris, France: http://basebe.obspm.fr".