Radiotherapy-Optimization/QP-Benchmark

QP-Benchmark

Benchmark instances for [Warm-IP](https://github.com/Radiotherapy-Optimization/Warm-IP), an open-source solver for large-scale convex quadratic programs (QPs), from the paper Warm-IP: A Path-Following ADMM Warm Start for Interior-Point Quadratic Programming (Aslani, Tefagh, Jhanwar, Zarepisheh; preprint link to be added). Every instance is a convex QP

minimize 0.5 x'Qx + q'x + c subject to constraints (one-sided or two-sided; see below)

stored as an HDF5 (.h5) file, one folder per family:

MPCdata/ 64 model-predictive-control instances MPC001....h5 MMdata/ 137 Maros-Meszaros instances MM001....h5 IMRTlungdata/ 60 radiotherapy (lung IMRT) instances IMRTlung001.h5

Radiotherapy data is organized one folder per (modality, site) family -- IMRT_lung today; future releases may add sibling families such as IMRT_prostate or VMAT_lung -- and radiotherapy instance names are exactly <family>_<index>. Everything else about an instance (patient, treatment protocol, provenance) lives in the manifest (instances_metadata.csv, one row per instance: family, name, source, sizes, nonzero counts, protocol, the spelled-out problem formulation, note) and in each file's data_note, never in folder or file names.

Using the data

The Warm-IP repository downloads these files automatically (its src/data_loader.py fetches any requested instance on first use), so nothing needs to be downloaded by hand. To fetch files directly:

from huggingface_hub import hf_hub_download
path = hf_hub_download(
    repo_id="Radiotherapy-Optimization/QP-Benchmark", repo_type="dataset",
    filename="IMRT_lung_data/IMRT_lung_001.h5")

File format

Q (n x n, full symmetric) and the constraint matrix G (m x n) are stored in CSR form as four datasets each (Q_data, Q_indices, Q_indptr, Q_shape, and likewise for G); q and c are stored directly. Two constraint forms are used:

• —Canonical (radiotherapy files) -- dataset h is present: G x <= h.
• —Two-sided (MM and MPC files) -- datasets l_h, u_h, l_x, u_x are present: l_h <= G x <= u_h, l_x <= x <= u_x, with infinite entries encoding absent bounds; a row with l_h[i] == u_h[i] is an equality constraint.

Every file carries format_note (format description) and data_note (provenance). See the data documentation in the Warm-IP repository for full details, family descriptions, and references.

Sources

• —MPC: the qpbenchmark MPC test set (Caron et al., 2024), qpsolvers/mpc_qpbenchmark.
• —MM: the Maros-Meszaros convex QP test set (Maros and Meszaros, 1999), via qpsolvers/maros_meszaros_qpbenchmark.
• —IMRT_lung: fluence-map-optimization QPs built from the public lung-patient data of PortPy (Jhanwar et al., 2023), one instance per patient under the Lung_2Gy_30Fx protocol.

Citing

If you use this data, please cite the Warm-IP paper (BibTeX in the repository README) and the original sources above for the family you use.