A Reproducible Data-Processing and Evaluation Framework for Federated Agricultural Knowledge Repositories

Agricultural bibliographic catalogs are difficult to benchmark for distributed learning because the records remain at the centers that produced them. We process 75,897 records from the International System for Agricultural Science and Technology (AGRIS), held by seven area-prefix clients, into a federated multi-view benchmark with four disjoint views (1096 features) and ten AGROVOC classes. Publication-level grouping precedes a client- and class-stratified 70/10/20 split; without it, 1132 duplicate-document groups would have crossed folds. In a blinded 400-record audit, the heuristic agreed with the consensus labels for 94.50% of the records (macro-F1 0.945; sampling-weighted accuracy 94.91%). On one shared schedule and a pre-specified budget of 40 rounds, federated averaging (FedAvg) reaches a macro-F1 of 0.2998±0.0151, below centralized training (0.4123±0.0066) and local-only training (0.4020±0.0175). Its validation curve has not converged by round 40, so the gap is a budget result, not a property of a fully trained model. FedProx and 8-bit quantization (0.3001±0.0147) stay within the seed-to-seed variation of FedAvg; quantization reduces uplink traffic 4-fold (total traffic 1.6-fold). The contribution is the benchmark and its evaluation protocol. Zenodo release v1.2 supports recomputation of the reported results and, when the AGRIS files match the stored checksums, reconstruction of the source text.