A metadata-first audit of a YOLO-format tomato leaf disease Kaggle dataset: ten classes, a 4.7-to-1 imbalance between the largest and smallest, a 16-image discrepancy between the metadata manifest and the local filesystem, and a training plan for the YOLOv8-s baseline that comes next.

A supervised model on a 528,000-row cleaned FAERS extract, reframed as a queue-ranking problem rather than a classifier. Random forest reaches PR-AUC 0.759 and recall 0.730 on hospitalization after six outcome-bearing fields are excluded to prevent leakage.

A bigram TF-IDF linear SVM on the Akbulut AI and Human text corpus reaches 0.9984 accuracy on a 1,214-row holdout. Exactly two predictions are wrong. One of them is a human writer labeled as AI. That cell is the whole story.

On a 44,898-article Kaggle fake news dataset, a TF-IDF classifier using only the subject and date fields reaches F1 = 1.000. The text-based headline model lands at 0.9980. That gap is the real finding.

A stroke screening classifier on a dataset where 95 percent of rows are negative. The 94 percent accurate model flagged almost nothing. The 75 percent accurate model caught four out of five cases. Here is why the second one was the right answer.

My read of a strange but serious Kawasaki disease paper, the North Pacific wind pattern behind it, and the pathway I would test next.

A practical look at the California housing prices dataset: what the rows actually mean, why the target is capped, and what this dataset is good for teaching.

Most explanations of neural networks start in the wrong place. This is the mental model that finally made layers, weights, activations, and backprop feel concrete to me.