Open-source alternatives to Constitutional Ai

Interpret is an interpretable machine learning library and glassbox model framework. It provides toolkits for training inherently transparent models and applying post-hoc explanation techniques to make machine learning predictions human-understandable. The framework distinguishes itself by integrating differential privacy into the training of interpretable models to prevent sensitive data from leaking through explanations. It also features a visualization tool for rendering interactive decision paths and model behavior. The library covers model explainability through feature importance calcu

Garak is a suite of tools for measuring AI reliability, scanning for vulnerabilities, and automating security assessments through adaptive probing. It functions as a generative AI vulnerability scanner and evaluation tool designed to identify security gaps, hallucinations, and failure modes in language models. The framework provides a toolkit for red-teaming and safety assessments, utilizing a structured system of probes and detectors to calculate failure rates. It specifically scans for risks such as data leakage and prompt injection by recording model responses to adversarial inputs. The p

This project is an agnostic model interpretability framework and explainability tool designed to provide local interpretable explanations for individual predictions. It functions as a local surrogate model that approximates the behavior of any machine learning classifier or regression model to identify the most influential features for a specific instance. The framework is designed to be model-agnostic, meaning it can explain predictions across tabular, text, and image data regardless of the underlying architecture. It employs local linear approximations and feature importance visualization t

NeMo-Guardrails is a toolkit for adding programmable safety constraints and dialogue boundaries to large language model conversational systems. It functions as security middleware that intercepts inputs and outputs to block prompt injections, jailbreaks, and sensitive data leaks, while providing a conversational dialogue manager to define structured interaction flows through configuration files. The framework includes a hallucination filter to screen model outputs for factual accuracy and a specialized modeling language for defining conversational flows and constraints. It provides capabiliti

The evaluation dataset data/samples-1680.jsonl.gz is the test set used in the following paper:

Status: Archive (code is provided as-is, no updates expected)

PySyft is a privacy-preserving machine learning framework and remote computation engine. It functions as a decentralized data analysis orchestrator that allows for the execution of data science workflows on remote servers without requiring the transfer of raw private data from the host device. The platform provides a secure collaboration environment where data owners manage permissions and authorize specific collaborators to run computations. It differentiates its workflow by utilizing mock data for local development and validation before submitting final analysis jobs to private remote serve

Source code/webpage/demos for the What-If Tool

LLM Guard is a security firewall and guardrail framework designed to scan and sanitize inputs and outputs for large language models. It functions as a proxy gateway and security layer to block prompt injections, toxicity, and sensitive data leakage while ensuring that model interactions remain compliant with organizational policies. The system distinguishes itself through a modular scanner pipeline that utilizes local model orchestration to eliminate external network dependencies. It supports real-time security filtering via streaming chunk analysis and implements a fail-fast execution model

LLM Prompt Injection Detector

Training PyTorch models with differential privacy

Algorithms for explaining machine learning models

SHAP is a machine learning explainer that uses a game-theoretic framework to estimate the contribution of each feature to a model prediction. It provides a set of tools for quantifying how individual input features push a specific output away from a baseline value. The project includes specialized explainers for different architectures, including high-speed implementations for decision trees and ensemble models, linearization algorithms for deep learning networks, and covariance integration for linear models. It also features a model-agnostic interpretability tool that uses a kernel method to

The Adversarial Robustness Toolbox (ART) is an open-source library that provides a unified framework for evaluating, defending, and certifying machine learning models against adversarial threats. It wraps models from any framework behind a common estimator interface, enabling composable pipelines for attack generation, defense application, robustness certification, and privacy auditing across evasion, poisoning, and extraction threats. The library distinguishes itself by covering the full adversarial ML security lifecycle within a single toolkit. It supports gradient-based adversarial example

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.

Updated the multilingual model weights used by Detoxify with a model trained on the translated data from the 2nd Jigsaw challenge (as well as the 1st). This model has also been trained to minimise bias and now returns the same categories as the unbiased model. New best AUC score on the test set:…