anonym-submit-26/bess-bench-26

BESS-Bench 是一个用于机器学习时间序列研究的恒星光谱数据集，专门针对Be星光谱。该数据集基于BeSS社区的光谱档案，提供了长期、多质量的光谱时间序列，用于表示学习和时间建模。数据集包含339,115个光谱行（Parquet行），对应37,624个物理观测（每个观测会话）和1,468颗独特恒星，时间跨度35年（1990年3月至2025年12月）。其中91.6%的观测来自业余观测者。数据集提供全波段光学光谱，波长覆盖范围为1150–10442 Å，并包含多条谱线（如Hα、Hβ、He I 5876等）的覆盖统计。数据集附带预训练的Hα窗口光谱编码器（80.3万参数）和三个基准下游任务：SpecProbe（Hα单谱线特征回归）、LineTransfer（Hβ到Hα的跨线泛化）和EWForecast（短时标EW(Hα)预测）。数据集采用按星分裂（训练/验证/测试），避免时间泄漏，并提供了详细的模式，包括波长、流量、流量误差、恒星名称、坐标、信噪比、观测日期、仪器设置等列。数据集基于CC-BY-4.0许可，要求引用原始BeSS数据库。BESS-Bench is a dataset of stellar spectroscopic time series for machine learning research, specifically focused on Be-star spectra. It provides long-baseline, mixed-quality spectra from the BeSS community archive for representation learning and temporal modeling. The dataset contains 339,115 spectral rows (parquet rows), corresponding to 37,624 physical observations (one observing session) and 1,468 unique stars, spanning 35 years (March 1990 to December 2025). Amateur observers contribute 91.6% of the observations. It offers full-range optical spectra with native wavelength coverage from 1,150 to 10,442 Å, including coverage statistics for multiple spectral lines (e.g., Hα, Hβ, He I 5876). The dataset comes with a pre-trained Hα-window spectral encoder (803K parameters) and three benchmarked downstream tasks: SpecProbe (Hα single-line spectral-feature regression), LineTransfer (Hβ → Hα cross-line generalization), and EWForecast (short-horizon EW(Hα) forecasting). It uses a per-star split (train/validation/test) to prevent temporal leakage and includes a detailed schema with columns such as wavelength, flux, flux_error, star_name, coordinates, SNR, observation date, instrument setup, etc. The dataset is licensed under CC-BY-4.0 and requires attribution to the original BeSS database.