RecourseBench documentation

RecourseBench is a modular framework for reproducible algorithmic recourse evaluation. It composes five extensible component types — datasets, preprocessing steps, target models, recourse methods, and evaluation metrics — into experiments that are configured as data and run end to end.

Live example

Edit the factual case and select a recourse method to see the kind of result RecourseBench evaluates: a counterfactual that changes the model prediction with a small set of feature edits.

Toy credit recourse run

Editable applicant, same toy model, different recourse method.

Select method Wachter DiCE Growing Spheres

Factual case

Prediction

Income ($k) Savings ($k) Debt ratio (%) Late payments

Apply edits and check

Counterfactual result

validity

L1 distance

features changed

Getting started

Install the package, run your first experiment from a YAML config, and use the Python API.

Getting started

API reference

Full description of every public class and function: arguments, return values, and behaviour.

API reference

Methods

Browse all registered recourse methods and open the RecourseBench paper.

Methods

Extending the framework

Register new datasets, preprocessing, models, methods, and metrics.

Extending the framework

Agent tools

Use the RecourseBench skills and MCP server from coding agents without giving up reproducibility or safety.

Agent tools