A comprehensive set of fairness metrics for datasets and machine learning models, explanations for these metrics, and algorithms to mitigate bias in datasets and models.

WEFE: The Word Embeddings Fairness Evaluation Framework. WEFE is a framework that standardizes the bias measurement and mitigation in Word Embeddings models. Please feel welcome to open an issue in case you have any questions or a pull request if you want to contribute to the project!

Toolkit for Auditing and Mitigating Bias and Fairness of Machine Learning Systems 🔎🤖🧰

NeurIPS 2019 Paper: RUBi : Reducing Unimodal Biases for Visual Question Answering

[ICML 2022] Channel Importance Matters in Few-shot Image Classification

Estimation and inference from generalized linear models using explicit and implicit methods for bias reduction

Methods for M-estimation of statistical models

This repository contains the experiments conducted in the ICLR 2022 spotlight paper "On the Importance of Firth Bias Reduction in Few-Shot Classification".

Bias correction command-line tool for climatic research written in C++

Bias reduction in quasi likelihood estimation

Bias detection Toolkit: Chrome Extension, Python Package, SOTA research paper docs.

Tensorflow implementation of Learning Not to Learn (CVPR 2019)

This repository contains the firth bias reduction experiments on the few-shot distribution calibration method conducted in the ICLR 2022 spotlight paper "On the Importance of Firth Bias Reduction in Few-Shot Classification".

Sampling algorithms and machine learning models to reduce bias and predict credit risk.

Pytorch implementation of 'Explaining text classifiers with counterfactual representations' (Lemberger & Saillenfest, 2024), ECAI 2024 - 27th European conference on AI

Apply empirical bias-reduced methods to fit a variety of latent variable models

A small and simple prototype designed to alert users of the bias of the news source.

Nexus.ai is a secure, vendor-neutral AI orchestration engine. It lets multiple LLMs and web search debate a question, then ranks and reconciles outputs to reduce bias and surface the best supported answer (with citations/media). All I/O is encrypted (AES-256), every step is logged for auditability, and the project ships without API keys by default.

A method to preprocess the training data, producing an adjusted dataset that is independent of the group variable with minimum information loss.

unbiased toxicity detection from comments