sba-scoring-v2
Native CatBoost binary classifier, selected campaign model mixed-v1-0051.
Positive class 1 means CHGOFF. Frozen weights, sigmoid calibration and operating
threshold 0.167204479067272 are preserved without refitting.
Nine inputs use the existing sba-features-1 / sba-scoring-1 contract.
Frozen inference files are in model/; supporting service, contracts, dependency
locks and standalone inference command are in service/.
Manifest SHA-256: 5d5498e8d3339ffbbce3d5f117f52692cee9986067a3bf6df8524c5e4c8c53ac.
Parameter-driven loading
SCORING_ENABLED=true
SCORING_MODEL_PATH=hf://thedauns/sba-scoring-v2@<immutable-40-character-HF-commit-SHA>/model
HF_CACHE_DIR=/tmp/huggingface
HF_LOCAL_FILES_ONLY=false
HF_HUB_OFFLINE=0
Replace the placeholder with the immutable SHA from this repository's commit
history. Local directory paths are also supported. Only runtime files download;
the service checks the pinned manifest and all file hashes before inference.
For private repos supply HF_TOKEN as an environment secret.
Standalone inference (Python 3.11, no E5 model or service key required):
cd service
python -m pip install -r requirements-scoring.txt
SCORING_MODEL_PATH=../model python -m scripts.predict_scoring --request examples/scoring-request.json
To download remotely, set SCORING_MODEL_PATH to the pinned HF URI above.
The HTTP service also requires the separate E5 model configured as documented in
the application. See service/SCORING_HF.md for cache, Docker and offline details.
Historical validation and limits
FY2023: precision 58.958130%, recall 70.081019%, F1 64.040190%;
TP 1211 / FP 843 / FN 517 / TN 7392, 9963 completed loans.
This cohort was inspected during model selection and is historical diagnostic
validation, not a fresh independent test. The 85% recall / 55% precision target was
not reached. See model/README.md, training.json and evaluation.json for split
provenance, technical release gates and later-year degradation.
This is a US SBA to RU demonstration, not production banking approval. Historical TermInMonths availability is unconfirmed. Java decision policy is unchanged. No raw loan dataset, research joblib or other model weights are included.