Responsible AI toolkit to build reliable and fair models
Model governancein 3 simple steps
With a model card as the source of truth, you can easily align across stakeholders on model details, considerations, parameters, performance explainability and fairness results.
Gather inputs and align stakeholders across product, data science and compliance
Source of truth for model details, performance, explainability and fairness analysis
Reports and Checks
Automated documentation of model and alerts from performance / fairness checks.
Simple yet comprehensive
Integrate your favourite data science models with 5 types of tests covering more than 20 different types of fairness requirements.
Improve model reliability
Have the confidence and peace of mind that models being deployed meets prescribed performance and fairness criteria.
Satisfies compliance requirements
Aligns with both the spirit and requirements of Singapore's MAS fairness assessment methodology.
Model Testing for Data Teams
Avoid unintended biases from development to productionRead the Docs
Subgroup Disparity Test
Assert that the model's performance across subgroups is similar.
Min / Max Metric Threshold Test
Ensure a minimum level of acceptable service across subgroups of interest.
Test the robustness of the model by perturbing protected variables of interest.
Feature Importance Test
Check if certain protected attributes are in the model's most influential features.
Data Shift Test
Assert that the data distribution is similar between 2 datasets (e.g. development and production).
Looking for other methods?
Open a new Github issue and let us know the types of model tests that you would like to see implemented!
From Development to Production
Write data science code as tests and produce model reports for documentation, accountability and reporting purposes
Get started with VerifyML by bootstrapping a model card using our web tool, going through our getting started notebook, or viewing the details of a model card with our online model card explorer.