awslabs/fortuna

Provides framework-agnostic calibration and conformal prediction methods for classification and regression, plus Bayesian inference procedures for Flax models to quantify both aleatoric and epistemic uncertainty. Operates across three usage modes: post-hoc calibration of pre-trained model outputs, conformal prediction from existing uncertainty estimates, and end-to-end Bayesian training with JAX/Flax. Returns rigorous prediction sets with guaranteed coverage guarantees rather than point estimates, enabling safety-critical deployments.

921 stars. No commits in the last 6 months.