foundry-tabular-credit-g

Binary classifier trained on Statlog German credit (UCI). Source rows 1000.

Data

• Dataset license: CC BY 4.0, verified on the UCI Machine Learning Repository page on 2026-10-06. Weights are a derivative of that data and are released under the same license; credit the UCI Machine Learning Repository and the dataset's authors. Code files (treepack.py, trees.mojo) are MIT.
• Split: 60/20/20 stratified, seed 42. Selection on validation only.
• Sensitive attributes present in features: age, personal_status, foreign_worker. No fairness evaluation was performed.

Metrics (ROC AUC)

• Validation: 0.8294
• Held-out test (sklearn original): 0.7986
• Held-out test (exported model, this file): 0.7986
• Max |probability difference| to sklearn on test rows: 2.22e-16

Use and limits

• Research baseline on one public dataset. Not for credit, hiring, insurance or medical decisions.
• Distribution shift, calibration and subgroup behaviour are untested.
• model.npz contains arrays only (no pickle). SHA-256 7b6e55b4a8b44e7cb33aa614a0e72c7595f6afea162d1ee4e9ca93b8e790a910, 5370 bytes.
• Inference: from treepack import Forest; Forest("model.npz").predict_proba(df). Uses a Mojo kernel when libfoundry_trees.so is built, else a numpy fallback.

CPU throughput (2 cores, Mojo kernel, includes preprocessing)

batch	sklearn rows/s	exported rows/s	ms/call
1	97	180	5.545
64	5814	11369	5.629
1024	71248	153234	6.683
8192	218364	748252	10.948