delta-audit

Delta-Audit Documentation

Welcome to the Delta-Audit documentation! Delta-Audit is a lightweight Δ-Attribution suite for auditing model updates (A→B) with behavioural linkage and robustness checks.

What is Δ-Attribution?

Δ-Attribution (Delta-Attribution) is a framework for understanding how model explanations change when models are updated. When you train a new version of a model, not only do the predictions change, but also the explanations of those predictions. Delta-Audit provides metrics to quantify and analyze these changes.

Key Concepts

Quick Start

# Install Delta-Audit
pip install delta-audit

# Run a quick demonstration
delta-audit quickstart

# Run the full benchmark
delta-audit run --config configs/full_benchmark.yaml

Core Metrics

Delta-Audit implements several key metrics:

Supported Algorithms

Supported Datasets

Documentation Sections

Citation

If you use Delta-Audit in your research, please cite:

@article{hemmat2025delta,
  title={Delta-Audit: Explaining What Changes When Models Change},
  author={Hemmat, Arshia},
  journal={arXiv preprint},
  year={2025}
}

Getting Help