What it does

This repo is a set of Jupyter notebooks that apply conformal prediction to off-the-shelf ML outputs. You get prediction sets or intervals with mathematically guaranteed coverage—90% of true labels fall inside, 90% of true values land in the interval—without touching the underlying model weights. The notebooks auto-download precomputed outputs and sample data, so you can experiment in a sandbox without wrangling ResNet training pipelines or multi-gigabyte medical datasets.

The interesting bit

The sandbox design is the quiet killer feature. Raw model outputs and data subsamples are fetched automatically; you need about 1.5GB of disk space and a “run all cells” reflex. Each notebook also carries a Google Colab link in its first cell, which lowers the activation energy from “interesting paper” to “actually running code” to roughly zero.

Key highlights

• Eight worked examples spanning ImageNet classification, medical expenditure regression, MS-COCO multilabel, toxic text outlier detection, tumor segmentation, and weather time-series under distribution shift
• Precomputed outputs from models like ResNet152 and Gradient Boosting Regressors—no GPU hours required
• Methods include conformalized quantile regression, weighted conformal prediction for drift, and selective classification that abstains when uncertain
• Accompanies the arXiv paper A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification
• Video tutorial trilogy embedded in the README for the visually inclined

Caveats

• The repo is explicitly educational scaffolding, not a pip-installable library; you’ll be copy-pasting notebook logic into your own pipeline
• Dependency management is conda-only; the README warns you to activate the conformal environment inside Jupyter or face the consequences

Verdict

Grab this if you keep reading about conformal prediction and need to see it actually work on real data before you trust it. Skip it if you’re hunting for a production-grade Python package with clean APIs—this is a curriculum, not a framework.