Browse Teach Curate

Greedy Decoding and Its Degeneration into Repetition

always taking the argmax is the obvious move and it produces loops — students need to feel that failure before sampling makes sense

jeremy Video

LLM Decoding Strategies: Greedy Search, Top-K, Nucleus Sampling, and Min-P in Machine Learning

Decoding strategies constitute a family of stochastic and deterministic mechanisms within machine learning that map probability distributions over vocabulary tokens to discrete sequence generations. …