An open benchmark for deep learning–based event reconstruction in neutrino telescopes — seven large-scale simulated datasets across six detector geometries, with baseline results for four reconstruction models on five core tasks.
Baseline performance of four reconstruction algorithms on the NuBench test partitions, exactly as reported in the paper. Pick a task, compare models across geometries, then drill into per-dataset metrics.
Simulated νμ charged- and neutral-current interactions produced with Prometheus, with a simplified detector response. Geometries are inspired by — but not identical to — existing and proposed telescopes in water and ice. Each dataset ships in SQLite and Parquet, with train/test selections, baseline predictions, and trained model artifacts.
Two architectures in active use within existing collaborations, one graph transformer, and one Kaggle-winning transformer — all implemented and trained in the open-source GraphNeT framework.
Grab the format you prefer from the dataset cards above — SQLite and Parquet contain identical data.
wget -P data/ \
<download-link>
All files end in .tar.gz, but only some are actually
compressed.
# SQLite & model artifacts tar -xzf file.tar.gz # Parquet & predictions tar -xf file.tar.gz
Both formats are compatible with the GraphNeT Dataset classes (SQLite recommended). Train/test partitions are defined by the included selection files.
from graphnet.data.dataset import (
SQLiteDataset,
)
If you use the NuBench datasets, predictions, or model artifacts in your work, please cite:
@article{orsoe2025nubench,
title = {NuBench: An Open Benchmark for Deep Learning--Based Event
Reconstruction in Neutrino Telescopes},
author = {{\O}rs{\o}e, Rasmus F. and Meighen-Berger, Stephan and Lazar,
Jeffrey and Prado, Jorge and Moz{\'u}n-Mateo, Iv{\'a}n and
Rosted, Aske and Weigel, Philip and Llorente Anaya, Arturo},
journal = {arXiv preprint arXiv:2511.13111},
year = {2025}
}