orailix/ride-gold-lite

RIDE Gold Lite

RIDE Gold Lite is the smaller benchmark-ready release of the RIDE dataset. It contains fixed train/test snapshot splits, a canonical evaluation table, and model-ready representations for train delay prediction on Belgian passenger railway operations.

This release mirrors the structure and prediction task of RIDE Gold Standard, but uses fewer snapshots and rows for faster inspection, development, and lower-cost experimentation.

Links

• Paper: https://arxiv.org/abs/2606.05070
• Code repository: https://github.com/orailix/ride

Files

Path	Description
core/dataset_core_spec.yaml	Benchmark split and target construction specification.
core/metadata.yaml	Core dataset metadata, parameters, and row counts.
core/test_eval_table.parquet	Canonical test evaluation table with snapshot keys, last known delay, future event metadata, and future delay targets. This is the main parquet table exposed in the Dataset Viewer.
tabular/	Dense tabular NumPy arrays for train and test splits, plus schema and normalization files. Used by MLP, XGBoost, and Transformer baselines.
sequential/	Sequential NumPy arrays for train and test splits, plus schema and normalization files. Used by the LSTM baseline.
gnn/	PyTorch Geometric heterogeneous graph shards for train and test splits, plus feature specification and normalization files. Used by the GNN baseline.
graph_event/	Graph-event benchmark artifacts, including topology links, test events, test journeys, and travel-time samples. Used by the deterministic graph-event baseline.
ride_gold_lite_croissant.json	Croissant metadata with schema, Responsible AI, and provenance information.

Coverage

• Train snapshot period: January 2023 to December 2024
• Test snapshot period: January 2025 to December 2025
• Training snapshots: 15,000
• Test snapshots: 3,000
• Future target events per train: 15
• Test evaluation rows: 623K
• Model-ready representations: tabular, sequential, GNN, and graph-event

Loading Example

import numpy as np
import pandas as pd

test_eval = pd.read_parquet("core/test_eval_table.parquet")

x_train = np.load("tabular/train/x.npy")
y_train = np.load("tabular/train/y.npy")
y_train_mask = np.load("tabular/train/y_mask.npy")
md_train = np.load("tabular/train/md.npy", allow_pickle=True)

For complete loading, training, and evaluation examples for the model-specific representations in this repository, see the code repository linked above.

Construction

RIDE Gold Lite was generated from RIDE Silver with the RIDE gold-construction pipeline. See the code repository linked above for the construction code, benchmark scripts, and configuration files.

Notes

RIDE Gold Lite contains derived benchmark artifacts, model arrays, graph tensors, operational railway records, and weather-derived variables. It does not intentionally contain passenger identities, employee identities, ticketing records, device identifiers, or direct personal information.

Responsible AI, provenance, and limitation metadata are provided in ride_gold_lite_croissant.json.

Sources

RIDE Gold Lite is derived from RIDE Silver:

• RIDE Silver: https://huggingface.co/datasets/orailix/ride-silver

The underlying railway and weather sources are:

• Infrabel Open Data: https://infrabel.opendatasoft.com/pages/home/
• Open-Meteo: https://open-meteo.com/

When using RIDE Gold Lite, please attribute RIDE, Infrabel Open Data, and Open-Meteo.

License

RIDE Gold Lite is released under CC BY 4.0.

Citation

@misc{elliker2026rideopendatasetbenchmark,
      title={RIDE: An Open Dataset and Benchmark for Train Delay Prediction},
      author={Clément Elliker and Mathis Le Bail and Clément Mantoux and Jesse Read and Sonia Vanier},
      year={2026},
      eprint={2606.05070},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2606.05070},
}