📝HardNLP FundamentalsPREMIUM

Decoding Strategies: Greedy to Nucleus

Master decoding strategies for text generation: compare greedy, beam search, top-k, nucleus (top-p), and min-p sampling, with temperature scaling and repetition penalty.

What you'll master

Greedy decoding as argmax: deterministic but degenerate

Beam search with length penalty: structured tasks only

Temperature scaling: controls distribution sharpness mathematically

Top-k sampling: fixed vocabulary truncation

Top-p (nucleus) sampling: dynamic cumulative probability threshold

Min-p sampling: confidence-scaled threshold, robust at high temperature

Top-p failure at high temperatures (long tail problem)

Repetition and frequency/presence penalties

Production pipeline order: repetition penalty → temperature → truncation

Human text has high perplexity: maximum likelihood is un-human

Hard30 min readIncludes code examples, architecture diagrams, and expert-level follow-up questions.

Premium Content

Unlock the full breakdown with architecture diagrams, model answers, rubric scoring, and follow-up analysis.

Code examplesArchitecture diagramsModel answersScoring rubricCommon pitfallsFollow-up Q&A

Want the Full Breakdown?

Premium includes detailed model answers, architecture diagrams, scoring rubrics, and 66 additional articles.

Back to Topics