lemma-baseline

name	lemma-baseline
description	Establish a dumb baseline and an honest validation harness before any real model, so every later number means something.
homepage	https://github.com/tkpratardan/lemma
license	MIT

Baseline

Deliver

Provide a reproducible validation design, a no-information or rule baseline, the simplest credible model, and the precise score later models must beat.

Check

• Define the outcome, observation unit, prediction moment, intended use, and a fixed seed.
• Match the split to deployment: time-ordered for temporal prediction, grouped for repeated entities, otherwise a fixed random split.
• Keep preprocessing inside the training pipeline.
• Choose a decision-relevant metric; do not rely on accuracy alone for imbalanced outcomes.
• Compare subgroup behavior and variation across folds or seeds when material.

Do not tune complex models, repeatedly inspect a final test set, preprocess before splitting, or celebrate a score without the dumb baseline.

For feature iteration and metric guidance, read references/deep-guide.md.