conformal-prediction

Why we included this project

Conformal prediction turns a model's raw scores into prediction sets or intervals whose coverage guarantees hold without assumptions about the data distribution, and this repo is one of the more approachable places to see it working on real problems. Each notebook pairs the method with a trained model: Imagenet classification with ResNet152, medical expenditure regression, MS-COCO multilabel recognition, gut-polyp segmentation, and weather forecasting under distribution shift, so you can compare how the same technique behaves across vision, NLP, and time-series data. The practical detail is that the notebooks pull precomputed model outputs and small data samples automatically, which means you can test and adapt conformal methods without re-running heavy models or downloading large datasets. That makes it useful both as a template for teams that want calibrated uncertainty around an existing model and as an executable tutorial for people new to the topic.