Bean Shape Lab model card
A seven-class multinomial logistic regression trained from scratch on UCI Dry Bean morphological measurements, with training-only standardization. The JSON artifact contains coefficients, intercepts, scaler statistics, class order and feature order; it is not a pickle. Use the accompanying repository for inference.
Data and license
Koklu and Özkan, Dry Bean (2020), UCI Machine Learning Repository, DOI 10.24432/C50S4B, dataset record. Dataset license: CC BY 4.0. Credit the dataset authors and cite the dataset and accompanying paper when reusing it. The original data are not redistributed here. The model and code are MIT licensed; this does not change the dataset license.
Evaluation
Stratified random 60/20/20 train/validation/test split, seed 42, with 8,166 / 2,722 /
2,723 rows. Fixed C=1 and no tuning. Validation accuracy 0.9243; held-out test accuracy
0.9210 and macro F1 0.9346; training-majority baseline accuracy 0.2607. The test
confusion matrix and per-class results are in metrics.json.
Notably, exact feature vectors appear across split boundaries: 19 distinct vectors in both train and test. This can inflate the reported score. A random split is not a farm-, lot-, camera-, or time-held-out evaluation. Other measurement methods and conditions may perform worse. The model is an educational research demo, not a production seed sorter; no calibration, out-of-distribution detector, image pipeline or operational guarantees are provided. Probabilities are not guaranteed calibrated.
Run provenance, dataset checksum, model checksum, split and training commit are in
training-manifest.json; local execution is CPU-only. No personal data are used.