This repository provides access to datasets, model predictions, and model artifacts for the NuBench dataset catalogue presented in NuBench: An Open Benchmark for Deep Learning–Based Event Reconstruction in Neutrino Telescopes. Datasets are available in two formats (SQLite and Parquet) - users may choose their preferred format.
The catalogue contains seven datasets simulated across six distinct detector geometries that are inspired by, but not strictly identical to, existing or proposed neutrino telescopes.
| Dataset | SQLite | Parquet | Predictions | Model Artifacts |
|---|---|---|---|---|
| Cluster | Download | Download | Download | Download |
| Flower L | Download | Download | Download | Download |
| Flower S | Download | Download | Download | Download |
| Flower XL | Download | Download | Download | Download |
| Hexagon | Download | Download | Download | Download |
| Hexagon Ice LE | Download | Download | Download | Download |
| Triangle | Download | Download | Download | Download |
- SQLite/Parquet: Dataset files in your preferred format. Both formats contain the same data, and are compatible with the Dataset classes in GraphNeT, which you may reuse in your own experiments. We recommend the SQLite Dataset Class. Test/train partitions are defined in the selection files.
- Predictions: Model predictions on the test-partition of the dataset from ParticleNet, DynEdge, GRIT, DeepIce.
- Model Artifacts: Trained model files and associated artifacts for reproducibility. This includes weights, pickled models and configuration files that allow you to re-create the models in GraphNeT.
To download the files using your terminal, copy the download link from the table above and run e.g. :
wget -P your_local_directory download_link
All files end in .tar.gz but only some have been subject to compression.
Extract the contents of model artifacts and SQLite downloads using
tar -xzf filename.tar.gz
The contents of predictions and Parquet downloads should be extracted using
tar -xf filename.tar.gz
NuBench - https://arxiv.org/pdf/2511.13111 Prometheus - https://arxiv.org/pdf/2304.14526 GraphNeT - https://arxiv.org/abs/2501.03817