DeciShift Digits XGBoost Baseline & Candidate Models
Deterministic XGBoost checkpoints used by the public DeciShift decision-change case study.
This model repository contains the two trained model artifacts that DeciShift compares inside a larger executable decision system. DeciShift itself is a behavioral-regression testing toolkit, not a single pretrained model; these checkpoints are published separately so the benchmark model layer is inspectable and reproducible.
Files
baseline.json— XGBoost booster trained for 20 boosting rounds (digits_xgb_v1).candidate.json— XGBoost booster trained for 25 boosting rounds (digits_xgb_v2).metadata.json— reproducibility metadata, evaluation metrics, source commit, and training configuration summary.
Task
Binary classification over the public scikit-learn Digits dataset: predict whether a handwritten digit is 8 (1) or not 8 (0).
The models operate on 64 pixel-intensity features. DeciShift's case study then combines model scores with versioned policy and rules components to produce final operational actions.
Evaluation snapshot
Using the deterministic held-out split from the DeciShift public example:
| Metric | Baseline | Candidate |
|---|---|---|
| Accuracy | 90.26% | 91.79% |
| ROC AUC | 0.9429 | 0.9472 |
These predictive metrics describe only the model layer. The associated DeciShift benchmark demonstrates that improved model metrics can still coincide with changed downstream decisions once policy and rule versions are included.
Reproduce
Generated from DeciShift source commit 4f0608cc38485e4a622e7f3b4623e38ecf5644b3 with:
python -m pip install -e '.[dev,xgboost]'
python huggingface/build_model.py
Load with XGBoost
import xgboost as xgb
baseline = xgb.Booster()
baseline.load_model("baseline.json")
candidate = xgb.Booster()
candidate.load_model("candidate.json")
Inputs must use the same 64-feature ordering as the DeciShift example (pixel_00 through pixel_63) after the example's V1 feature scaling (value / 16.0).
Related artifacts
- Source: https://github.com/sauravsingla/DeciShift
- Hugging Face Space: https://huggingface.co/spaces/sauravsingla08/DeciShift
- Benchmark dataset: https://huggingface.co/datasets/sauravsingla08/DeciShift-Decision-Change-Benchmark
- PyPI: https://pypi.org/project/decishift/
Intended use
These checkpoints are intended for reproducibility, testing, education, and inspection of DeciShift's public decision-change benchmark. They are not presented as general-purpose digit-recognition models or as production recommendations.
Limitations
The benchmark uses a fixed public dataset split and a deliberately compact binary task. Results do not establish production safety, fairness, compliance, causal attribution, or fitness for deployment. Downstream decision behavior depends on the full executable system, including policy and rules components, not only these model files.
License
The generated DeciShift model artifacts are published under Apache-2.0. Upstream source-data terms for the original scikit-learn Digits dataset remain applicable to the source data.