|
Download README.md from rocky250/PACE: direct link, hf CLI and curl.
- Browser
- Download file 2.96 kB
-
https://huggingface.co/rocky250/PACE/resolve/main/README.md
- Command line
-
hf download hf://rocky250/PACE/README.md
-
curl -L -o README.md https://huggingface.co/rocky250/PACE/resolve/main/README.md
2.96 kB
| tags: | |
| - tabular-classification | |
| - sme | |
| - credit-risk | |
| - calibrated-classification | |
| library_name: scikit-learn | |
| # PACE-SME Proposed Model | |
| This repository contains the fitted PACE-SME proposed two-component cross-fitted simplex classifier, its inference helper, feature schema, runtime dependencies, and aggregated evaluation assets. | |
| ## Intended use | |
| Research-only retrospective classification against the `target_missed_payment` label in the supplied Bangladesh SME dataset. The current data provenance and prediction-time horizon are not fully verified, so this artifact must not be represented as a validated prospective early-warning system. | |
| The model returns the positive-class score and a threshold-based label. It is a research screening signal only. A qualified human reviewer must assess the original records and context. Do not use this score as the sole basis for credit approval, denial, pricing, restructuring, or another consequential borrower decision. It does not estimate causal restructuring effects or identify an optimal intervention. | |
| ## Model details | |
| - Architecture: two-component cross-fitted probability simplex. | |
| - Components selected using development-only grouped out-of-fold results: CatBoost_Unweighted, LightGBM_Unweighted. | |
| - Calibration: isotonic. | |
| - Validation-selected decision threshold: 0.22682374. | |
| - Input features: see `feature_schema.json` and `model_config.json`. | |
| - Training and locked-test rows: 960 and 240. | |
| - Current nested-CV status: COMPLETED. | |
| ## Inference | |
| Install `requirements.txt`, then load the model with: | |
| ```python | |
| import json | |
| from pace_inference import load_predictor | |
| predictor = load_predictor('.') | |
| with open('example_input.json', encoding='utf-8') as input_file: | |
| record = json.load(input_file) | |
| result = predictor.predict(record) | |
| print(result) | |
| ``` | |
| Every feature in `feature_schema.json` is required. Extra input columns are ignored. Missing feature columns raise an error. Feature values must follow the schema and preprocessing learned during training. Replace each null placeholder in `example_input.json` with a valid value before prediction. | |
| ## Evaluation assets | |
| Aggregated validation and locked-test metrics are in `performance/metrics.json`; comparison tables are in the same directory; generated ROC, precision-recall, calibration, confusion-matrix, decision-curve, threshold-sensitivity, and capacity plots are in `figures/`. Metrics are produced by the current notebook run. They must be interpreted with their dataset, split, and evidence limitations. | |
| ## Reproducibility and limitations | |
| The fitted artifact is serialized with cloudpickle and depends on the exact package versions recorded in `requirements.txt`. Do not load pickle-based model files from untrusted sources. Raw datasets, row-level records, and row-level test predictions are not included in this repository. Dataset license and redistribution permissions must be checked before making this repository public. | |