Soybean Grain RGB Dataset and Group-Disjoint Benchmark Artifacts (48,973 grains, two variants)

RGB image dataset of individual soybean grains for classification research, released with the benchmark artifacts of the IEEE Access article "Partitioning Protocol Governs Reported Performance in a Single-Setup Soybean Grain Dataset: An Open RGB Pipeline and a Group-Disjoint Benchmark". Contents dataset_com_fundo.zip — 48,973 grain crops (background preserved), 8 classes, PNG. dataset_sem_fundo.zip — 48,432 grain crops (background removed), produced by a separate segmentation pass with its own bounding boxes, 8 classes, PNG. splits_agrupados_10sementes.csv — photograph-level partition manifest of the group-disjoint protocol for the ten seeds used in the article. manifests_random.zip — per-grain manifests of the random-protocol arm (ten seeds over the full 48,973-instance tree, as in the recorded runs; memberships reproduce the original per-class test supports exactly). manifests_random_limpos.zip — per-grain manifests of the filtered random arm re-run (ten seeds over the 48,039-instance tree), regenerated deterministically with the engine's stratified split and validated exactly against the re-run's per-instance predictions. resultados_por_execucao.zip — per-run metrics, per-instance predictions, runtime reports, and configurations for the reported experiment blocks. resultados_aleatorios_por_execucao.zip — per-run metrics, runtime reports, and configurations of the random per-grain arm's recorded runs (this arm has no per-instance predictions). resultados_aleatorios_limpos_por_execucao.zip — per-run metrics, per-instance predictions, runtime reports, and configurations of the random per-grain arm re-run on the 48,039-instance filtered tree (6 architectures × 10 seeds); supersedes the earlier random-arm runs for the article's protocol contrast. resultados_vazamento_por_execucao.zip— per-run metrics, per-instance predictions, runtime reports, and configurations of the fixed-test leakage-isolation experiment (clean/leaked training pairs, 2 architectures × 10 seeds; manifests in manifests_bloco8.zip). manifests_bloco8.zip — training manifests (clean/leaked pairs, 10 seeds) of the leakage-isolation experiment. CHECKSUMS_dataset.txt — SHA-256 of every package. Classes Instances in the background-preserved variant: Normal 20,236; Impurities 14,348; Broken 6,584; Shriveled 5,422; Green 1,192; Castor bean 553; Purple 398; Insect-damaged 240 — derived from source photographs captured under a single acquisition setup. Grains from one photograph share acquisition conditions, which is why the article's benchmark partitions at the photograph level. Crops are stored as 224x224 PNG (grain centered, area interpolation). Directory names (Portuguese): RGB_JPG_Normais = Normal; RGB_JPG_IMPUREZAS = Impurities; RGB_JPG_QUEBRADOS = Broken; RGB_JPG_CHOCHOS = Shriveled; RGB_JPG_ESVERDEADO = Green; RGB_JPG_PURPURAS = Purple; RGB_JPG_MAMONAS = Castor bean; RGB_JPG_INSETOS = Insect-damaged. The _SF suffix marks the background-removed variant. Known label ambiguity 934 records from photographs DSC_0320–DSC_0323 appear under both Shriveled and Normal (467 in each class). The article's experiments exclude all of them and use the remaining 48,039 instances from 466 single-class photographs; users of this dataset should handle these records accordingly. Code https://github.com/thesis-ufu/soybean-grain-classification — training pipeline, configs, partition manifests, and analysis scripts (uv-managed environment, Python 3.11). The repository's scripts/reproduce.py regenerates the article's tables directly from resultados_por_execucao.zip. License Dataset under CC BY 4.0; code under MIT (in the repository). If you use this dataset, please cite the article above (article DOI to be added upon publication). The article's code snapshot is archived at 10.5281/zenodo.21987916.