sauravsingla08/DeciShift-Decision-Change-Benchmark

DeciShift Decision-Change Benchmark

Record-level evidence for analyzing how final ML-system decisions change between baseline and candidate software versions.

This dataset is generated reproducibly from the public DeciShift repository. The first configuration uses scikit-learn's bundled Digits dataset with trained XGBoost models and versioned policy/rules components. It is designed to make decision-level behavioral regression inspectable rather than to provide another copy of the underlying image dataset.

Why this dataset exists

Model-level metrics can improve while operational decisions still move. In the committed Digits/XGBoost case study:

• evaluation records: 719
• baseline accuracy: 90.26%
• candidate accuracy: 91.79%
• baseline ROC AUC: 0.9429
• candidate ROC AUC: 0.9472
• changed final actions: 30 (4.17%)
• Decision Contract result: BLOCK

The global shift can stay within its declared limit while a governed cohort exceeds its own limit. That is the kind of regression DeciShift is built to surface.

Canonical Dataset Viewer split

The Hugging Face Dataset Viewer is explicitly configured to load only:

data/digits_xgboost.parquet

The changed-record convenience artifact and aggregate summaries remain downloadable evidence, but they are intentionally excluded from the canonical split so the Viewer cannot duplicate records or infer an incompatible schema from auxiliary files.

Files

• data/digits_xgboost.parquet — canonical 719-record dataset split with baseline/candidate actions, transition labels, model scores, cohort context, contract context, and attribution-share context.
• analysis/digits_xgboost_changed.parquet — convenience artifact containing changed-decision records only; it is intentionally excluded from the canonical dataset split to avoid duplicate rows in the Dataset Viewer.
• summaries/digits_xgboost.json — machine-readable aggregate case-study evidence.
• summaries/scenarios.jsonl — feature/model/policy/rules/all-change scenario summaries.

Core columns

Column	Meaning
record_id	Stable evaluation-record identifier for this deterministic split
actual_digit	Ground-truth digit label from the public Digits dataset
baseline_score / candidate_score	Baseline/candidate model score when emitted by the flow
baseline_action / candidate_action	Final operational action from each executable system version
changed	Whether the final action changed
transition	Explicit action transition, or unchanged
cohort	Governed cohort label used for slice analysis
cohort_shift_rate	Observed action-shift rate for that cohort
cohort_contract_limit	Declared cohort limit used by this example
cohort_contract_violation	Whether the observed cohort rate exceeds its declared limit
global_contract_result	PASS/BLOCK-style result produced by the Decision Contract evaluation
changed_nodes	Versioned nodes that differ between baseline and candidate
attribution_share_*	Global absolute software-counterfactual attribution share for changed nodes
source_commit	Git commit that generated the artifact

Reproducibility

Generated from DeciShift source commit 4f0608cc38485e4a622e7f3b4623e38ecf5644b3 by:

python huggingface/build_dataset.py

Related artifacts

• Interactive Space: https://huggingface.co/spaces/sauravsingla08/DeciShift
• Source repository: https://github.com/sauravsingla/DeciShift
• PyPI: https://pypi.org/project/decishift/

Important limits

This is descriptive evaluation evidence over a fixed public split. Software-counterfactual attribution does not establish real-world causality. A Decision Contract pass or block only reflects user-declared limits over the supplied evidence; it is not proof of safety, fairness, compliance, correctness, or production fitness.

The repository does not republish the raw 8×8 pixel arrays. Source dataset provenance is retained through the upstream scikit-learn Digits reference; upstream dataset terms continue to apply to source data.

License

DeciShift-generated code and dataset artifacts in this repository are published under Apache-2.0. Upstream source-data terms remain applicable to the original source dataset.