When NumPy broadcasting gives you the wrong loss
A shape-compatible operation can still compare the wrong values. Diagnose a silent pairwise-error bug before it reaches your metrics.
2026-10-21
THE REFERENCE SHELF
Guides
Explanations you can work with.
Split first, fit second: a preprocessing leakage check
Use a small scaler diagnostic to see whether held-out data is influencing training, then put the boundary inside a scikit-learn pipeline.
2026-10-21
Estimate KV-cache memory before choosing a context length
Calculate the dense cache payload from layers, KV heads, token count and dtype, and see why the answer is not a GPU-fit guarantee.
2026-10-21