NuBench

NuBench is an open benchmark dataset for the development and evaluation of reconstruction algorithms for Cherenkov neutrino telescopes, analogous in spirit to PILArNet for LArTPC detectors. It provides a common, well-documented simulation framework that allows the community to train and compare machine learning models on a shared dataset without requiring access to proprietary experimental data.

The dataset represents neutrino events in a large-volume Cherenkov detector through detector hit records in multiple complementary representations: graph and point-cloud formats suited to geometric deep learning (e.g. graph neural networks), and tabular formats for classical methods. Data are stored in Parquet and SQLite formats and hosted on ERDA (Electronic Research Data Archive).

NuBench is designed around a set of benchmark reconstruction tasks that map onto the core challenges in Cherenkov neutrino telescope physics: event classification, energy reconstruction, direction reconstruction, and particle identification. Reference model implementations based on GraphNeT — a framework for graph neural network-based reconstruction in neutrino telescopes — are provided on GitHub together with data loaders and training scripts, enabling reproducible comparisons across methods.

Documentation is available via the dataset paper on arXiv and supplementary material on GitHub.