Part 3: The Science - Hyperparameter Tuning & Getting to 100% Precision with Warp/Oz

Part 2 ended with a dataset: 973 annotated 10-second chunks from 15 roasting sessions, recording-level splits, class-weighted training. What it didn't cover is what happened when that dataset went into a training loop for the first time, and why the first result (91.1% accuracy, 87.5% precision, 3 false positives) was unacceptable for an automated roasting assistant. This post is about the two things that got the model to 97.4% accuracy, 100% precision, and 0 false positives . Both are domain de