Inspect the systems that make model results meaningful
A model evaluation can be misleading because its report is wrong. Historical training features can contain information the live predictor never had. A retrieval pipeline can discard useful results while merging ranked sources. nrml’s ML and AI exercises focus on these engineering failures in small, inspectable repositories.
You work with supplied data and code, rather than training a large model or calling an external model service to solve the exercise. The challenge is to explain and repair the behavior while preserving the application workflows around it.
A progression from metrics to historical data
Begin with the classifier evaluation report and calculate a tiny cohort by hand. Move to hybrid retrieval to reason about ranked evidence and deterministic output. Use point-in-time training features when you want a deeper exercise in availability, revisions, and historical reconstruction.
Before editing, state the evaluated population or time boundary in plain language. After editing, demonstrate a case that distinguishes the original failure from the intended behavior. The accompanying guide walks through independent examples you can use to practice that explanation.