BACI-VI-Bench: A dataset of variational inequality benchmark instances for multi-agent trade-network equilibrium Authors: Pham Thanh Hieu, Nguyen Kieu Linh Published: 2026 ABSTRACT This article presents BACI-VI-Bench, a processed benchmark dataset and reproducible construction pipeline that transforms CEPII-BACI product-level international trade records into finite-dimensional variational inequality instances for multi-commodity trade-network equilibrium. The source data are CEPII-BACI HS17 bilateral trade files, version V202601, covering annual exporter-importer-product flows from 2017 to 2024. The pipeline filters positive bilateral trade flows, maps six-digit Harmonized System products to commodity sectors, selects leading exporting and importing economies, aggregates values and quantities into structured flow tensors, normalizes observed trade volumes, and calibrates a Nagurney-style benchmark cost-and-price operator. The dataset provides year-level instances of dimension 500 (five commodity sectors, ten exporters, ten importers, one route) and sector-level instances of dimension 100 for five HS commodity groups: Machinery, Minerals, Chemicals, Transport Equipment, and Metals. Each variational inequality instance contains a feasible set, a calibrated operator, observed and normalized flow tensors, exporter and importer identifiers, sector labels, operator parameters, and projection residual diagnostics. The repository includes Python construction and validation scripts, benchmark characterization data for extragradient and self-adaptive inertial solvers, figures, and metadata files enabling reuse with projection, extragradient, self-adaptive variational inequality algorithms, and multi-agent reinforcement learning frameworks. The benchmark dataset was originally developed in the context of a broader study on trade-network equilibrium modeling. Nevertheless, the present data article provides a complete and self-contained description of the dataset, benchmark construction pipeline, instance format, metadata, validation procedures, and reproducibility resources, enabling independent reuse of the benchmark. Specifications TableSubjectComputer Sciences; Applied Mathematics; Economics and EconometricsSpecific subject areaVariational inequality benchmark data for trade-network equilibrium and multi-agent reinforcement learningType of dataProcessed data; filtered data; analyzed data; tables; compressed arrays (NPZ); figures; Python scripts; JSON; CSV; Markdown documentationData formatNPZ (NumPy compressed); CSV; JSON; PNG; Python (.py); Markdown (.md)Data collectionCEPII-BACI HS17 V202601 annual CSV files (approximately 10–11.5 million bilateral HS6-digit trade records per year, covering up to 226 reporting countries and 5381 product codes at the HS6 level; 8 annual files for 2017–2024) were processed with Python. Bilateral trade flows were filtered, aggregated from HS6 products into five commodity sectors via HS2 chapter mapping, normalized, and converted into finite-dimensional variational inequality instances with calibrated operators and feasible sets. The top 10 exporters and top 10 importers by aggregate trade value were retained, yielding year-level networks of dimension d = 500 (K = 5 sectors, m = 10 exporters, n = 10 importers) and sector-level networks of dimension d = 100. The processed benchmark comprises 13 NPZ instances (8 year-level, 5 sector-level), with observed flow vectors retaining approximately 92% of possible bilateral-sector pairs on average.Data source locationInstitution: Posts and Telecommunications Institute of Technology (PTIT)City: HanoiCountry: VietnamLatitude and longitude: Not applicable (processed benchmark derived from international trade records)Data accessibilityRepository name: ZenodoData identification number: https://doi.org/10.5281/zenodo.20607238 (Version v0.3)Direct URL to data: https://zenodo.org/records/20607238Code repository: https://github.com/hieuthanhpham12/BACI-VI-BenchInstructions for accessing: Publicly accessible without restriction. Download processed NPZ and CSV files and run validate_baci_vi_bench.py. Raw BACI source files must be obtained separately from CEPII (www.cepii.fr) [1]. The benchmark dataset and code are available on Zenodo [10].Related research articleP.T. Hieu and N.K. Linh, MARL-Guided Self-Adaptive Inertial Extragradient Methods for Variational Inequality Models of Multi-Commodity Trade Network Equilibrium, submitted to Computational and Applied Mathematics, 2026 [2]. SubjectComputer Sciences; Applied Mathematics; Economics and EconometricsSpecific subject areaVariational inequality benchmark data for trade-network equilibrium and multi-agent reinforcement learningType of dataProcessed data; filtered data; analyzed data; tables; compressed arrays (NPZ); figures; Python scripts; JSON; CSV; Markdown documentationData formatNPZ (NumPy compressed); CSV; JSON; PNG; Python (.py); Markdown (.md)Data collectionCEPII-BACI HS17 V202601 annual CSV files (approximately 10–11.5 million bilateral HS6-digit trade records per year, covering up to 226 reporting countries and 5381 product codes at the HS6 level; 8 annual files for 2017–2024) were processed with Python. Bilateral trade flows were filtered, aggregated from HS6 products into five commodity sectors via HS2 chapter mapping, normalized, and converted into finite-dimensional variational inequality instances with calibrated operators and feasible sets. The top 10 exporters and top 10 importers by aggregate trade value were retained, yielding year-level networks of dimension d = 500 (K = 5 sectors, m = 10 exporters, n = 10 importers) and sector-level networks of dimension d = 100. The processed benchmark comprises 13 NPZ instances (8 year-level, 5 sector-level), with observed flow vectors retaining approximately 92% of possible bilateral-sector pairs on average.Data source locationInstitution: Posts and Telecommunications Institute of Technology (PTIT)City: HanoiCountry: VietnamLatitude and longitude: Not applicable (processed benchmark derived from international trade records)Data accessibilityRepository name: ZenodoData identification number: https://doi.org/10.5281/zenodo.20607238 (Version v0.3)Direct URL to data: https://zenodo.org/records/20607238Code repository: https://github.com/hieuthanhpham12/BACI-VI-BenchInstructions for accessing: Publicly accessible without restriction. Download processed NPZ and CSV files and run validate_baci_vi_bench.py. Raw BACI source files must be obtained separately from CEPII (www.cepii.fr) [1]. The benchmark dataset and code are available on Zenodo [10].Related research articleP.T. Hieu and N.K. Linh, MARL-Guided Self-Adaptive Inertial Extragradient Methods for Variational Inequality Models of Multi-Commodity Trade Network Equilibrium, submitted to Computational and Applied Mathematics, 2026 [2]. Value of the Data •BACI-VI-Bench provides empirically grounded variational inequality instances derived from real bilateral trade flows, enabling algorithm testing on structured real-world network data. Researchers in variational inequality theory, network optimization, and multi-agent systems can use these instances to benchmark projection, extragradient, inertial, self-adaptive, and fixed-point-constrained algorithms without collecting or preprocessing raw trade data.•Each NPZ file contains calibrated operators, exporter-wise feasible sets, observed and normalized flow tensors, operator parameters, and projection residual diagnostics in a single self-contained file. These data allow direct algorithm evaluation and comparison using a common, fully documented instance format.•The eight year-level instances (2017–2024, d=500) and five sector-level instances (d=100) provide a structured family of instances at two controlled scales. The year-level instances support longitudinal comparisons, while the sector-level instances support controlled single-commodity experiments.•The same variational inequality instances can be reinterpreted as multi-agent trade-network environments, with exporters and importers as agents, flow vectors as states, and natural projection residuals as equilibrium proximity measures. This dual interpretation makes the data useful for research at the intersection of variational inequalities and multi-agent reinforcement learning.•Python construction and validation scripts, metadata tables, and benchmark characterization data are provided for transparency and reproducibility. The pipeline can be re-run on later BACI releases to generate updated instances, extending the dataset's utility beyond the V202601 version.•Unlike many existing variational inequality test instances that rely on synthetic operators and randomly chosen constraint sets, BACI-VI-Bench provides operators and feasible sets calibrated from recorded bilateral trade flows. This grounding in structured real-world data allows researchers to evaluate algorithm behavior on economically meaningful problem instances. The data-calibrated components - observed flow tensors, exporter-wise capacity constraints, and unit-value-based operator coefficients (aik, dik, pjk) - are derived directly from recorded BACI bilateral trade data, whereas the congestion slope b = 0.25 and transport-cost loading τ = 0.05 are fixed benchmark defaults documented in metadata/benchmark_config_default.json.•The dataset is distributed as a standalone benchmark resource, including benchmark instances, metadata, validation scripts, and characterization outputs, allowing researchers to reproduce, compare, and extend computational studies independently of any particular algorithmic implementation. •BACI-VI-Bench provides empirically grounded variational inequality instances derived from real bilateral trade flows, enabling algorithm testing on structured real-world network data. Researchers in variational inequality theory, network optimization, and multi-agent systems can use these instances to benchmark projection, extragradient, inertial, self-adaptive, and fixed-point-constrained algorithms without collecting or preprocessing raw trade data. BACI-VI-Bench provides empirically grounded variational inequality instances derived from real bilateral trade flows, enabling algorithm testing on structured real-world network data. Researchers in variational inequality theory, network optimization, and multi-agent systems can use these instances to benchmark projection, extragradient, inertial, self-adaptive, and fixed-point-constrained algorithms without collecting or preprocessing raw trade data. •Each NPZ file contains calibrated operators, exporter-wise feasible sets, observed and normalized flow tensors, operator parameters, and projection residual diagnostics in a single self-contained file. These data allow direct algorithm evaluation and comparison using a common, fully documented instance format. Each NPZ file contains calibrated operators, exporter-wise feasible sets, observed and normalized flow tensors, operator parameters, and projection residual diagnostics in a single self-contained file. These data allow direct algorithm evaluation and comparison using a common, fully documented instance format. •The eight year-level instances (2017–2024, d=500) and five sector-level instances (d=100) provide a structured family of instances at two controlled scales. The year-level instances support longitudinal comparisons, while the sector-level instances support controlled single-commodity experiments. The eight year-level instances (2017–2024, d=500) and five sector-level instances (d=100) provide a structured family of instances at two controlled scales. The year-level instances support longitudinal comparisons, while the sector-level instances support controlled single-commodity experiments. •The same variational inequality instances can be reinterpreted as multi-agent trade-network environments, with exporters and importers as agents, flow vectors as states, and natural projection residuals as equilibrium proximity measures. This dual interpretation makes the data useful for research at the intersection of variational inequalities and multi-agent reinforcement learning. The same variational inequality instances can be reinterpreted as multi-agent trade-network environments, with exporters and importers as agents, flow vectors as states, and natural projection residuals as equilibrium proximity measures. This dual interpretation makes the data useful for research at the intersection of variational inequalities and multi-agent reinforcement learning. •Python construction and validation scripts, metadata tables, and benchmark characterization data are provided for transparency and reproducibility. The pipeline can be re-run on later BACI releases to generate updated instances, extending the dataset's utility beyond the V202601 version. Python construction and validation scripts, metadata tables, and benchmark characterization data are provided for transparency and reproducibility. The pipeline can be re-run on later BACI releases to generate updated instances, extending the dataset's utility beyond the V202601 version. •Unlike many existing variational inequality test instances that rely on synthetic operators and randomly chosen constraint sets, BACI-VI-Bench provides operators and feasible sets calibrated from recorded bilateral trade flows. This grounding in structured real-world data allows researchers to evaluate algorithm behavior on economically meaningful problem instances. The data-calibrated components - observed flow tensors, exporter-wise capacity constraints, and unit-value-based operator coefficients (aik, dik, pjk) - are derived directly from recorded BACI bilateral trade data, whereas the congestion slope b = 0.25 and transport-cost loading τ = 0.05 are fixed benchmark defaults documented in metadata/benchmark_config_default.json. Unlike many existing variational inequality test instances that rely on synthetic operators and randomly chosen constraint sets, BACI-VI-Bench provides operators and feasible sets calibrated from recorded bilateral trade flows. This grounding in structured real-world data allows researchers to evaluate algorithm behavior on economically meaningful problem instances. The data-calibrated components - observed flow tensors, exporter-wise capacity constraints, and unit-value-based operator coefficients (aik, dik, pjk) - are derived directly from recorded BACI bilateral trade data, whereas the congestion slope b = 0.25 and transport-cost loading τ = 0.05 are fixed benchmark defaults documented in metadata/benchmark_config_default.json. •The dataset is distributed as a standalone benchmark resource, including benchmark instances, metadata, validation scripts, and characterization outputs, allowing researchers to reproduce, compare, and extend computational studies independently of any particular algorithmic implementation. The dataset is distributed as a standalone benchmark resource, including benchmark instances, metadata, validation scripts, and characterization outputs, allowing researchers to reproduce, compare, and extend computational studies independently of any particular algorithmic implementation. Background Variational inequalities provide a unifying mathematical framework or equilibrium problems in optimization, network economics, and multi-agent systems [3,4]. International trade networks can be represented as commodity flows between supply and demand nodes subject to feasibility constraints and equilibrium conditions, making them natural instances for variational inequality formulations [5]. A companion research article [2] introduced a self-adaptive inertial extragradient method for trade-network equilibrium computation and conducted experiments on BACI-derived network instances, motivating the need for a standalone, reusable benchmark dataset. The CEPII-BACI database provides harmonized bilateral trade flows at the HS 6-digit product level for a comprehensive set of countries and years [6]. BACI-VI-Bench transforms product-level BACI records into finite-dimensional variational inequality instances with explicit feasible sets, calibrated operators, residual diagnostics, and solver-ready files. The benchmark construction is independent of any specific algorithm: the dataset is designed for reuse across projection, extragradient, inertial, self-adaptive, and multi-agent methods. The MARL-VIP connection interprets equilibrium-seeking agents as solvers, with trade-flow adjustment policies corresponding to VI projection steps, providing a structured empirical environment for reinforcement learning research in network economics. Data Description The dataset is organized in a structured directory hierarchy to facilitate access, reuse, and reproducibility. The root of the Zenodo archive is the folder BACI-VI-Bench_Zenodo_v0.2/, which contains processed benchmark instances, benchmark characterization outputs, figures, construction and validation scripts, metadata files, and source documentation. The main directory structure is shown in Fig. 1. Table 1 lists all folders and major files with their formats, contents, and reuse purposes.Fig. 1Directory structure of the BACI-VI-Bench dataset, showing the processed instance folders (year-level and sector-level NPZ files), benchmark characterization outputs, figures, construction and validation scripts, metadata files, and raw source instructions. NPZ Instance Structure.Fig 1 dummy alt textTable 1Repository structure and contents of the BACI-VI-Bench dataset.Each NPZ file stores a single benchmark instance. Table 2 lists all arrays and their descriptions as defined in metadata/instance_schema.json.Table 1 dummy alt textFolder / FileFormatContentsReuseprocessed_instances/ year_instances/NPZEight compressed instances (2017–2024), one per year, dimension d=500(K=5 sectors, m=10 exporters, n=10 importers, L=1 route)VI algorithm input; MARL environmentprocessed_instances/ sector_instances/NPZFive sector-level instances for 2022 (Machinery, Minerals, Chemicals, Transport, Metals), dimension d=100Single-sector VI experimentsbenchmark_outputs/ benchmark_BACI_HS17_all_years.csvCSVBenchmark characterization data for four solvers across all year-level instances: iteration counts, final residuals, CPU timesInstance difficulty characterizationbenchmark_outputs/ benchmark_BACI_HS17_Y2022_sector.csvCSVBenchmark characterization data for five sector-level instances, year 2022Sector-level instance characterizationbenchmark_outputs/ benchmark_BACI_HS17_Y2022_ablation.csvCSVAblation diagnostic data for the 2022 full instance: comparison of SAISE variants with and without inertial extrapolation (θ=0vs.θ=0.6) alongside an EG reference; includes a prior SAISE parameter configuration for robustnessComponent-level instance characterizationbenchmark_outputs/ benchmark_BACI_HS17_Y2022_tuning.csvCSVHyperparameter sensitivity diagnostic data for the 2022 full instanceParameter sensitivity characterizationfigures/PNGFive publication-quality figures: convergence curves, flow heatmaps, network graph, multi-year comparison, sector comparison (Fig1-Fig5)Visual characterization of instancescode/build_baci_vi_bench.pyPythonConstruction script: reads raw BACI CSV files, writes NPZ instancesDataset regeneration from BACI sourcecode/trade_vi_instance.pyPythonModule: TradeVIInstance class with operator, projection, and residual methodsAlgorithm implementation scaffoldcode/validate_baci_vi_bench.pyPythonValidation script: checks file integrity, dimensions, feasibility, residualsAutomated integrity verificationcode/saise_baci_experiment.pyPythonFull original experiment script used to generate all benchmark characterization dataFull reproducibility of characterization datametadata/instance_schema.jsonJSONSchema of NPZ variable names, shapes, units, and descriptionsData documentation and interoperabilitymetadata/benchmark_config_default.jsonJSONDefault network configuration: years, dimensions, operator parameter valuesReproducible configuration referencemetadata/hs2_to_sector_mapping.csvCSVMapping from HS2 chapter codes (01-97) to commodity sector namesReproducible sector aggregationmetadata/file_index_sha256.csvCSVSHA-256 checksums for all generated data filesFile integrity verificationraw_source_manifest/MD, CSVInstructions for obtaining BACI source files from CEPII; BACI variable descriptions; license summaryProvenance and reproducibility documentationREADME.mdMarkdownDataset overview, VI formulation, file naming convention, how to generate and validate instances; minimal working exampleEntry point for new usersCITATION.cffCFFMachine-readable citation metadata for the datasetCitation automationLICENSE_CODE_MIT.txtTextMIT license covering the construction and validation codeReuse conditionsTable 2Variables stored in each NPZ instance. Tensor shape notation: K= sectors, m = exporters,n = importers, L = routes.Table 2 dummy alt textVariableShapeDescriptionx_obs(K,m,n,L)Normalized observed trade-flow tensor derived from BACI value column vflow_value_musd(K,m,n,L)Retained bilateral trade value in million USDflow_quantity_ton(K,m,n,L)Retained bilateral trade quantity in tonnes (where available in BACI)supply_cap(m,)Exporter-wise capacity vector defining the feasible set Cdemand_total(n,)Importer-wise observed demand totals in normalized unitsa_cost(K,m)Calibrated production-cost coefficients aikb_congestionscalarCongestion coefficient b (default value: 0.25)tau_transportscalarTransport-cost loading coefficient τ (default value: 0.05)p_price(K,n)Calibrated demand-price scale parameters pjkd_demand(K,n)Calibrated demand-sensitivity parameters djkuv_exporter_norm(K,m)Exporter-side normalized unit values used in operator calibrationuv_importer_norm(K,n)Importer-side normalized unit values used in operator calibrationnormalization_scale_musd(1,)Maximum retained flow value (million USD) used for normalizationexporter_codes(m,)BACI exporter country codesimporter_codes(n,)BACI importer country codesexporter_names(m,)Short exporter labelsimporter_names(n,)Short importer labelssector_names(K,)Commodity-sector labels derived from HS2 chapter groupingmetadata_json(1,) string arrayEmbedded JSON: year, BACI version, dimensions, filtering rules Directory structure of the BACI-VI-Bench dataset, showing the processed instance folders (year-level and sector-level NPZ files), benchmark characterization outputs, figures, construction and validation scripts, metadata files, and raw source instructions. NPZ Instance Structure. Table 1Repository structure and contents of the BACI-VI-Bench dataset.Each NPZ file stores a single benchmark instance. Table 2 lists all arrays and their descriptions as defined in metadata/instance_schema.json.Table 1 dummy alt textFolder / FileFormatContentsReuseprocessed_instances/ year_instances/NPZEight compressed instances (2017–2024), one per year, dimension d=500(K=5 sectors, m=10 exporters, n=10 importers, L=1 route)VI algorithm input; MARL environmentprocessed_instances/ sector_instances/NPZFive sector-level instances for 2022 (Machinery, Minerals, Chemicals, Transport, Metals), dimension d=100Single-sector VI experimentsbenchmark_outputs/ benchmark_BACI_HS17_all_years.csvCSVBenchmark characterization data for four solvers across all year-level instances: iteration counts, final residuals, CPU timesInstance difficulty characterizationbenchmark_outputs/ benchmark_BACI_HS17_Y2022_sector.csvCSVBenchmark characterization data for five sector-level instances, year 2022Sector-level instance characterizationbenchmark_outputs/ benchmark_BACI_HS17_Y2022_ablation.csvCSVAblation diagnostic data for the 2022 full instance: comparison of SAISE variants with and without inertial extrapolation (θ=0vs.θ=0.6) alongside an EG reference; includes a prior SAISE parameter configuration for robustnessComponent-level instance characterizationbenchmark_outputs/ benchmark_BACI_HS17_Y2022_tuning.csvCSVHyperparameter sensitivity diagnostic data for the 2022 full instanceParameter sensitivity characterizationfigures/PNGFive publication-quality figures: convergence curves, flow heatmaps, network graph, multi-year comparison, sector comparison (Fig1-Fig5)Visual characterization of instancescode/build_baci_vi_bench.pyPythonConstruction script: reads raw BACI CSV files, writes NPZ instancesDataset regeneration from BACI sourcecode/trade_vi_instance.pyPythonModule: TradeVIInstance class with operator, projection, and residual methodsAlgorithm implementation scaffoldcode/validate_baci_vi_bench.pyPythonValidation script: checks file integrity, dimensions, feasibility, residualsAutomated integrity verificationcode/saise_baci_experiment.pyPythonFull original experiment script used to generate all benchmark characterization dataFull reproducibility of characterization datametadata/instance_schema.jsonJSONSchema of NPZ variable names, shapes, units, and descriptionsData documentation and interoperabilitymetadata/benchmark_config_default.jsonJSONDefault network configuration: years, dimensions, operator parameter valuesReproducible configuration referencemetadata/hs2_to_sector_mapping.csvCSVMapping from HS2 chapter codes (01-97) to commodity sector namesReproducible sector aggregationmetadata/file_index_sha256.csvCSVSHA-256 checksums for all generated data filesFile integrity verificationraw_source_manifest/MD, CSVInstructions for obtaining BACI source files from CEPII; BACI variable descriptions; license summaryProvenance and reproducibility documentationREADME.mdMarkdownDataset overview, VI formulation, file naming convention, how to generate and validate instances; minimal working exampleEntry point for new usersCITATION.cffCFFMachine-readable citation metadata for the datasetCitation automationLICENSE_CODE_MIT.txtTextMIT license covering the construction and validation codeReuse conditions Repository structure and contents of the BACI-VI-Bench dataset. Each NPZ file stores a single benchmark instance. Table 2 lists all arrays and their descriptions as defined in metadata/instance_schema.json. Table 2Variables stored in each NPZ instance. Tensor shape notation: K= sectors, m = exporters,n = importers, L = routes.Table 2 dummy alt textVariableShapeDescriptionx_obs(K,m,n,L)Normalized observed trade-flow tensor derived from BACI value column vflow_value_musd(K,m,n,L)Retained bilateral trade value in million USDflow_quantity_ton(K,m,n,L)Retained bilateral trade quantity in tonnes (where available in BACI)supply_cap(m,)Exporter-wise capacity vector defining the feasible set Cdemand_total(n,)Importer-wise observed demand totals in normalized unitsa_cost(K,m)Calibrated production-cost coefficients aikb_congestionscalarCongestion coefficient b (default value: 0.25)tau_transportscalarTransport-cost loading coefficient τ (default value: 0.05)p_price(K,n)Calibrated demand-price scale parameters pjkd_demand(K,n)Calibrated demand-sensitivity parameters djkuv_exporter_norm(K,m)Exporter-side normalized unit values used in operator calibrationuv_importer_norm(K,n)Importer-side normalized unit values used in operator calibrationnormalization_scale_musd(1,)Maximum retained flow value (million USD) used for normalizationexporter_codes(m,)BACI exporter country codesimporter_codes(n,)BACI importer country codesexporter_names(m,)Short exporter labelsimporter_names(n,)Short importer labelssector_names(K,)Commodity-sector labels derived from HS2 chapter groupingmetadata_json(1,) string arrayEmbedded JSON: year, BACI version, dimensions, filtering rules Variables stored in each NPZ instance. Tensor shape notation: K= sectors, m = exporters,n = importers, L = routes. Load NPZ example Each NPZ file can be loaded directly in Python as follows: Minimal working example To facilitate immediate use of the benchmark, the following minimal example demonstrates how to load a benchmark instance, evaluate the variational inequality operator, project onto the feasible set, and compute the natural projection residual. A more comprehensive example is provided in the repository documentation and the accompanying trade_vi_instance.py module. Benchmark characterization data The four CSV files in benchmark_outputs/ contain diagnostic data characterizing each instance’s computational difficulty. These data are generated by running solvers on each instance and recording convergence behavior. The all-years CSV (benchmark_BACI_HS17_all_years.csv) records for each solver–instance combination: year, method label, last iteration, iteration counts to convergence at thresholds 10⁻⁶ and 10⁻³ (DNF if not reached within 3000 iterations), final projection residual G(xk), a secondary distance diagnostic (FinalDist), CPU time in seconds, backtrack counts, function evaluation counts, step-size λ, and hyperparameter values (θ,σ,μ). The sector CSV records analogous columns for two solvers on sector-level instances. The tuning CSV records hyperparameter combinations explored during a grid search on the year 2022 instance together with a summary convergence score. Figures Five figures are provided in the figures/ folder, each in PNG format at publication resolution. A complete mapping between each figure and its source CSV or NPZ files is recorded in metadata/figure_table_source_mapping.csv and docs/figure_table_source_mapping.md:•Fig1_BACI_HS17_Y2022_convergence.png: projection residual G(xk) vs. iteration number for three solvers on the 2022 year-level instance (d=500).•Fig2_BACI_HS17_Y2022_flow_heatmaps.png: heat maps of the normalized flow tensor xobs for each of the five commodity sectors in the 2022 instance.•Fig3_BACI_HS17_Y2022_network_graph.png: directed network graph of the 2022 trade network with edge weights proportional to observed normalized flows.•Fig4_BACI_HS17_multiyear_comparison.png: bar chart of final projection residuals and iteration counts across eight year-level instances for all four solvers.•Fig5_BACI_HS17_Y2022_sector_speedup.png: bar chart of iteration counts for two solvers on the five sector-level instances. •Fig1_BACI_HS17_Y2022_convergence.png: projection residual G(xk) vs. iteration number for three solvers on the 2022 year-level instance (d=500). Fig1_BACI_HS17_Y2022_convergence.png: projection residual G(xk) vs. iteration number for three solvers on the 2022 year-level instance (d=500). •Fig2_BACI_HS17_Y2022_flow_heatmaps.png: heat maps of the normalized flow tensor xobs for each of the five commodity sectors in the 2022 instance. Fig2_BACI_HS17_Y2022_flow_heatmaps.png: heat maps of the normalized flow tensor xobs for each of the five commodity sectors in the 2022 instance. •Fig3_BACI_HS17_Y2022_network_graph.png: directed network graph of the 2022 trade network with edge weights proportional to observed normalized flows. Fig3_BACI_HS17_Y2022_network_graph.png: directed network graph of the 2022 trade network with edge weights proportional to observed normalized flows. •Fig4_BACI_HS17_multiyear_comparison.png: bar chart of final projection residuals and iteration counts across eight year-level instances for all four solvers. Fig4_BACI_HS17_multiyear_comparison.png: bar chart of final projection residuals and iteration counts across eight year-level instances for all four solvers. •Fig5_BACI_HS17_Y2022_sector_speedup.png: bar chart of iteration counts for two solvers on the five sector-level instances. Fig5_BACI_HS17_Y2022_sector_speedup.png: bar chart of iteration counts for two solvers on the five sector-level instances. Experimental Design, Materials and Methods Source data and inclusion criteria The raw data source is the CEPII-BACI HS17 bilateral trade database, version V202601 [1,6]. BACI provides annual bilateral trade flows at the HS 6-digit product level. The columns used are: t (year), i (exporter country code), j (importer country code), k (HS6 product code), v(trade value in thousand USD), and q (trade quantity in tonnes, where available). Only records with strictly positive trade value (v>0) and valid HS6 codes are retained. BACI raw files are not included in the Zenodo archive; the file raw_source_manifest/how_to_obtain_baci_hs17_v202601.md provides instructions for downloading from CEPII. Eight annual datasets are used: 2017, 2018, 2019, 2020, 2021, 2022, 2023, and 2024, with 2022 as the primary reference year. Network construction: country and sector selection For each annual dataset, the top m=10 exporting economies and top n=10 importing economies are selected by aggregate bilateral trade value. The top K=5 HS commodity sectors by total trade value are retained. The number of routes is fixed at L=1 per exporter-importer pair because BACI does not contain route-level transport data. The year-level network dimension is d=K×m×n×L=500. For sector-level instances, K=1and d=100. The set of selected exporting and importing economies is determined independently for each year by aggregate trade value, so the composition of the ten selected countries may differ across annual instances. In contrast, the five commodity sectors are defined by fixed HS2 chapter mappings (Machinery: chapters 84–85; Minerals: chapters 25–27; Chemicals: chapters 28–38; Transport Equipment: chapters 86–89; Base Metals: chapters 72–83), which are consistent across all year-level and sector-level instances. As a result, the eight year-level instances share a common sector structure but are not constrained to include the same set of trading economies. HS6-to-sector aggregation HS 6-digit product codes are mapped to commodity sectors via their HS2 chapter codes. The mapping is provided in metadata/hs2_to_sector_mapping.csv. The five commodity sector groups are: Machinery and Electrical Equipment (HS2 chapters 84–85), Mineral Products (chapters 25–27), Chemical Products (chapters 28–38), Transport Equipment (chapters 86–89), and Base Metals (chapters 72–83). Within each sector, trade values and quantities are aggregated across all retained HS6 codes for each exporter-importer pair. Flow-value processing and normalization Trade values in thousand USD are converted to million USD. The normalization scale is the maximum retained flow value across the full tensor. The normalized observed flow tensor is:xobs[k,i,j,l]=flow_value_musd[k,i,j,l]normalization_scale_musd. All entries of xobs lie in [0, 1]. The normalization scale is stored in each NPZ file under the key normalization_scale_musd. Feasible set and capacity construction The feasible set is the nonnegative exporter-wise capacity set:C={x≥0:∑k,j,ℓx[k,i,j,ℓ]≤supply_cap[i],i=1,…,m}. The capacity vector supply_cap[i] is set to 1.5 times the total normalized outgoing flow of exporter iin the observed data, making xobsstrictly interior to C. Variational inequality operator calibration The benchmark operator is a Nagurney-style trade-network excess-cost-minus-incentive operator [5]. For sector k, exporter i, importerj, and route l:Fk,i,j,ℓ(x)=aik(1+b∑jx[k,i,j,ℓ])+τaik−pjk1+djk∑ix[k,i,j,ℓ]·where ∑j(...) denotes summation over importers and ∑i(...) over exporters. Operator parameter calibration: aik is proportional to the exporter-sector normalized unit value; pjk is proportional to the importer-sector normalized unit value; djk is calibrated from importer demand totals; b=0.25 and τ=0.05 are fixed benchmark defaults recorded in metadata/benchmark_config_default.json. All calibrated parameters are stored in each NPZ file and verified in metadata/calibration_audit.csv, which records zero absolute error across all 13 instances. The exact calibration formulas implemented in code/build_baci_vi_bench.py are: a_cost[k,i] = 0.5 * clip(uv_exporter_norm[k,i], 0.2, 5.0); d_demand[k,j] = 0.4 / max(import_total[k,j], 0.01), where import_total[k,j] = sum_{i,l} x_obs[k,i,j,l]; and p_price[k,j] = max(avg_cost[k,j] * (1 + d_demand[k,j] * import_total[k,j]), 0.8 * avg_cost[k,j]), where avg_cost[k,j] is the mean production-plus-transport cost over active exporters to importer j in sector k. Operator properties Under the benchmark calibration, F is Lipschitz continuous on C and pseudomonotone on C in the sense of network equilibrium theory [5]. These properties guarantee existence of at least one solution by the Hartman–Stampacchia theorem [3]. Strict monotonicity need not hold in general. A high-accuracy approximate reference solution can be obtained by running the included solvers from xobs with tolerance 10−8; pre-computed reference solutions will be added in a future dataset release. Projection onto the feasible set Projection PC(z) onto C is performed in two steps: (a) truncate all negative entries to zero; (b) for each exporteri, if ∑k,j,ℓz[k,i,j,ℓ]>supply_cap[i], rescale outgoing flows proportionally to satisfy the capacity constraint exactly. This procedure is equivalent to the Euclidean projection when exporter constraint sets are independent. The implementation in code/trade_vi_instance.py is verified to be idempotent to machine precision. Residual and gap diagnostics The benchmark residual is the natural projection residual:G(x)=‖x−PC(x−F(x))‖,where ‖·‖is the Euclidean norm. G(x)=0 if and only if xsolves the variational inequality. For each instance, G(xobs) is computed and recorded as a baseline diagnostic. The convergence threshold for the characterization solvers is G(xk) < 10–6. Benchmark characterization data generation Four solvers are run on each instance to generate diagnostic data characterizing convergence behavior and instance difficulty. All solvers are run for a maximum of 3000 iterations. The Method column in the CSV output uses the standardized labels EG_spectral, EG_oracle, IEG, and SAISE. The four solvers are:•EG (spectral step size; CSV label: EG_spectral): the extragradient method [7] with a step size derived from a spectral estimate of the operator’s Lipschitz constant.•EG (oracle grid step size; CSV label: EG_oracle): the extragradient method with the step size selected from a coarse grid, reported as an oracle-step reference.•IEG (oracle step size; CSV label: IEG): the inertial extragradient method [8] with an oracle step size.•SAISE (self-adaptive step size, θ=0.6,σ=0.65,μ=0.6; CSV label: SAISE): the self-adaptive inertial subgradient extragradient method [2,8,9] with Armijo-type adaptive step size and inertial extrapolation, with hyperparameters selected via grid search on the 2022 instance. •EG (spectral step size; CSV label: EG_spectral): the extragradient method [7] with a step size derived from a spectral estimate of the operator’s Lipschitz constant. EG (spectral step size; CSV label: EG_spectral): the extragradient method [7] with a step size derived from a spectral estimate of the operator’s Lipschitz constant. •EG (oracle grid step size; CSV label: EG_oracle): the extragradient method with the step size selected from a coarse grid, reported as an oracle-step reference. EG (oracle grid step size; CSV label: EG_oracle): the extragradient method with the step size selected from a coarse grid, reported as an oracle-step reference. •IEG (oracle step size; CSV label: IEG): the inertial extragradient method [8] with an oracle step size. IEG (oracle step size; CSV label: IEG): the inertial extragradient method [8] with an oracle step size. •SAISE (self-adaptive step size, θ=0.6,σ=0.65,μ=0.6; CSV label: SAISE): the self-adaptive inertial subgradient extragradient method [2,8,9] with Armijo-type adaptive step size and inertial extrapolation, with hyperparameters selected via grid search on the 2022 instance. SAISE (self-adaptive step size, θ=0.6,σ=0.65,μ=0.6; CSV label: SAISE): the self-adaptive inertial subgradient extragradient method [2,8,9] with Armijo-type adaptive step size and inertial extrapolation, with hyperparameters selected via grid search on the 2022 instance. Convergence iteration counts, final residuals, CPU times, backtrack counts, function evaluation counts, step-size values, and hyperparameter records are stored in the four CSV files in benchmark_outputs/. An ablation diagnostic file (benchmark_BACI_HS17_Y2022_ablation.csv) records four rows using standardized labels: EG_oracle (EG reference), SAISE_prior (an earlier SAISE parameter set with σ=0.9, μ=0.7), SAISE_no_inertia (SAISE without inertial extrapolation, θ=0), and SAISE (the tuned configuration, θ=0.6,σ=0.65,μ=0.6), enabling component-level sensitivity analysis on the 2022 year-level instance. Software environment All data processing and characterization experiments were conducted in Python 3.10 with NumPy 1.24, Pandas 2.0, and Matplotlib 3.7 on a standard CPU workstation. The construction script, validation script, and solver module are provided in the code/ folder with a requirements.txt file. No GPU hardware is required. Reproducibility instructions To regenerate the full dataset from the raw BACI source files: To validate all generated NPZ and CSV files: The validation script checks file integrity, array shapes, finite values, nonnegativity, capacity feasibility, and projection idempotency. A summary is written to logs/validation_log.txt. Limitations BACI provides annual bilateral aggregate trade flows but does not contain route-level transport data. The benchmark therefore uses L=1 route per exporter-importer pair by default. Extending to multiple routes would require supplementary transport data not present in BACI. The variational inequality operator is calibrated from aggregated trade values and quantities. It serves as a benchmark construction, not a structurally estimated econometric model. Researchers requiring a structurally identified equilibrium model should treat this operator as an initialization or benchmark reference. Selection of the top ten exporters, top ten importers, and top five commodity sectors keeps the benchmark computationally tractable and yields instances of controlled size. However, the resulting benchmark should not be interpreted as a complete representation of the global trade system. Smaller economies, less-traded sectors, and long-tail trading relationships are intentionally excluded. Therefore, conclusions drawn from these benchmark instances should be interpreted with caution when extrapolating to the broader global trade system. The year-level instances are intended primarily for comparative benchmarking across annual trade environments rather than for strict longitudinal analysis on an identical network topology. Some BACI quantity records are missing or expressed in heterogeneous units; the benchmark prioritizes value-based flow construction. The dataset is derived from BACI HS17 V202601; later BACI releases may yield different numerical instances due to data revisions. The current version does not include pre-computed high-accuracy reference solutions. Users requiring solution accuracy metrics can generate approximate reference solutions by running the included solvers to a tight tolerance G(x)<10−8. Pre-computed reference solutions are planned for a future dataset release. Ethics Statement The authors confirm that they have read and comply with the ethical requirements for publication in Data in Brief. The present work does not involve human subjects, animal experiments, or data collected from social media platforms. The dataset is derived from publicly available international trade records provided by CEPII-BACI and contains country- and product-level trade aggregates rather than personal or identifiable data. No ethical approval was required. CRediT Author Statement Pham Thanh Hieu: Conceptualization, Methodology, Software, Formal analysis, Data curation, Validation, Visualization, Writing – original draft, Writing – review & editing; Nguyen Kieu Linh: Conceptualization, Supervision, Project administration, Writing – review & editing.