BACI-VI-Bench v0.3 is a processed benchmark dataset and reproducible construction pipeline for variational inequality (VI) and multi-agent reinforcement learning (MARL) studies of multi-commodity trade-network equilibrium. The dataset is derived from CEPII-BACI HS17 V202601 bilateral trade records for the years 2017–2024.
The archive provides:
8 year-level VI benchmark instances for 2017–2024, each with dimension d = 500 (K = 5 commodity sectors, m = 10 exporters, n = 10 importers, L = 1 route);
5 sector-level instances for the 2022 reference year, each with dimension d = 100;
observed and normalized trade-flow tensors;
exporter-wise feasible sets and Nagurney-style benchmark trade-network operators;
data-calibrated operator coefficients derived from unit-value and demand information, together with fixed benchmark defaults for congestion and transport loading;
projection residual diagnostics and benchmark characterization data for four solvers: EG_spectral, EG_oracle, IEG, and SAISE;
Python scripts for dataset construction, validation, and benchmark characterization;
metadata files, SHA-256 checksums, validation logs, calibration audit files, and figure/table source mappings;
publication-quality figures and documentation for reuse.
Version v0.3 improves reproducibility and transparency by adding calibration-audit metadata, figure/table source mapping, processed-instance manifests, sector-instance manifests, validation logs, and updated Zenodo/GitHub metadata.
This dataset complements a companion research article on self-adaptive inertial extragradient methods for multi-commodity trade-network equilibrium submitted to Computational and Applied Mathematics (2026). The benchmark instances are independent of any specific algorithm and can be reused with projection, extragradient, inertial, self-adaptive, stochastic, and MARL-based equilibrium-seeking methods.
Source data: CEPII-BACI HS17 V202601. Raw BACI files are not included in this archive; see raw_source_manifest/baci_source_manifest.csv for instructions on obtaining the source data from CEPII.