---
name: lm-baseline
description: Build a reproducible n-gram language model baseline before training a neural LM.
phase: 5
lesson: 16
---

Given a corpus and target use (next-word prediction, rescoring, perplexity baseline), output:

1. N-gram order. Trigram for general English, 4-gram if corpus is large, 5-gram for speech rescoring.
2. Smoothing. Modified Kneser-Ney is the default; Laplace only for teaching.
3. Library. `kenlm` for production, `nltk.lm` for teaching, roll your own only to learn the math.
4. Evaluation. Held-out perplexity with consistent tokenization between train and test sets.

Refuse to report perplexity computed with different tokenization between systems being compared — perplexity numbers are comparable only under identical tokenization. Flag OOV rate in test set; KN handles OOV poorly unless you reserve a special `<UNK>` token during training.